Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Screening_in_two-dimensional_gauge_theories.txt	creates a state which is a combination of states belonging to the Hilbert space, i.e. the integer- charged states. We conjecture that the same phenomenon occurs for dynamical fermions and is the reason of the incompatibility of our string tension calculations and Eq. (1.1). When only the integer part of the probe charge has a physical meaning then the effects of fractional charges are projected out, which explains the vanishing of the string tension for all real Q.
Processed_Screening_in_two-dimensional_gauge_theories.txt	state, which accounts for the difﬁculties in observing the string breaking in lattice QCD simulations which use solely this operator.
Processed_Screening_in_two-dimensional_gauge_theories.txt	Then we proposed two generalizations of Wilson loops for arbitrary real charges Q. We found that for both of them the string tension vanishes for any Q, which would be in disagreement with the prediction of Coleman et al. [9]. Based on the experience from the pure gauge theory, we suggested an explanation by pointing out that in the compact formulation the effects of fractional charges are projected out of the Hilbert space and only the integer charges have a physical meaning. The value for the string tension between integer charges predicted by Eq. (1.1) is indeed zero for all values of the fermion mass.
Processed_Screening_in_two-dimensional_gauge_theories.txt	This work was partially supported through NCN grant nr 2011/03/D/ST2/01932, and by Foun- dation for Polish Science MPD Programme co-ﬁnanced by the European Regional Development Fund, agreement no. MPD/2009/6. Numerical simulations were performed on deszno supercom- puter at the Faculty of Physics, Astronomy and Applied Computer Science, Jagiellonian University, Cracow.
Processed_Screening_in_two-dimensional_gauge_theories.txt	J. Boorstein, D. Kutasov, Nucl. Phys. B421 (1994) 263 [hep-th/9401044], D. Kutasov, A. Schwimmer, Nucl. Phys. B442 (1995) 447 [hep-th/9501024].
Processed_Screening_in_two-dimensional_gauge_theories.txt	F. Antonuccio, S. Pinsky, Phys. Lett. B439 (1998) 142 [hep-th/9805188], F. Antonuccio, H. Pauli, S. Pinsky, S. Tsujimaru, Phys. Rev. D58 (1998) 125006 [hep-th/9808120], D. Dorigoni, G. Veneziano, J. Wosiek, JHEP 1106 (2011) 051 [arXiv:1011.1200 [hep-th]].
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	This paper is concerned with a problem of robust ﬁltering for a ﬁnite-dimensional linear discrete time invariant system with two output signals, one of which is directly observed while the other has to be estimated. The system is assumed to be driven by a random dis- turbance produced from the Gaussian white noise sequence by an unknown shaping ﬁlter. The worst-case performance of an estimator is quantiﬁed by the maximum ratio of the root- mean-square (RMS) value of the estimation error to that of the disturbance over stationary Gaussian disturbances whose mean anisotropy is bounded from above by a given parameter a > 0. The mean anisotropy is a combined entropy theoretic measure of temporal coloured- ness and spatial “nonroundness” of a signal. We construct an a-anisotropic estimator which minimizes the worst-case error-to-noise RMS ratio. The estimator retains the general struc- ture of the Kalman ﬁlter, though with modiﬁed state-space matrices. Computing the latter is reduced to solving a set of two coupled algebraic Riccati equations and an equation in- volving the determinant of a matrix. In two limiting cases, where a = 0 or a , the a-anisotropic estimator leads to the standard steady-state Kalman ﬁlter or the H¥ -optimal estimator, respectively.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	We consider a robust ﬁltering problem for a ﬁnite-dimensional linear discrete time invariant (LDTI) system which generates two output signals. One of the signals is directly observed, while the other is unknown and has to be estimated by ﬁltering the observation through a causal LDTI estimator. The underlying system is driven by an external disturbance which is assumed to be a stationary Gaussian sequence. The latter is produced from a white noise sequence with zero mean and identity covariance matrix by an unknown LDTI shaping ﬁlter whose transfer function belongs to the Hardy space H2.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	these random sequences. The aim of the estimator is to minimize this inﬂuence. Suppose the choice of a noise shaping ﬁlter is at the disposal of a hypothetical opponent whose aim is to maximize the error-to-noise RMS ratio. If the set of strategies of this opponent is the whole space H2, then the worst damage to the estimator performance (in terms of the error-to-noise RMS ratio) which the opponent can achieve coincides with the H¥ -norm of the error operator relating the estimation error to the disturbance. This effect is quantitatively the same as in a qualitatively different situation where the estimator performance is measured by the ℓ2-gain of the error operator with respect to nonrandom disturbances in the form of arbitrary square summable sequences. The latter setting is studied, for example, in [15, 16, 20, 27, 28]. On the other hand, if the opponent is allowed to produce only Gaussian white noise disturbances with scalar covariance matrices (that is, diagonal matrices with all equal diagonal entries), then the above mentioned RMS ratio reduces to the scaled H2-norm 2/√m of the error operator, where m is the dimension of the disturbance.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	Consider an intermediate situation where the opponent is restricted to produce disturbances whose mean anisotropy is bounded from above by a given nonnegative parameter a which quantiﬁes the amount of uncertainty in the probability law of the noise. The mean anisotropy is a combined entropy theoretic measure of temporal colouredness (that is, predictability) and spatial “nonroundness” of a stationary Gaussian sequence. The corresponding set of noise shaping ﬁlters is a cone in H2. The worst-case performance of an estimator is then quantiﬁed by the a-anisotropic norm of the error operator deﬁned as the maximum error-to-noise RMS ratio with respect to the class of stationary Gaussian disturbances whose mean anisotropy does not exceed a.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	The present paper is concerned with an a-anisotropic optimal ﬁltering problem of ﬁnding an estimator so as to minimize the a-anisotropic norm of the error operator. We derive equations for such estimator assuming the existence of a saddle point in this stochastic minimax problem. The estimator retains the general structure of the Kalman ﬁlter, though with modiﬁed state- space matrices. Computing the latter is reduced to solving a set of two coupled discrete time algebraic Riccati equations (DAREs) and an equation involving the determinant of a matrix. In two limiting cases, where a = 0 and a , the a-anisotropic estimator becomes the standard steady-state Kalman ﬁlter [1] and the H¥ -optimal estimator, respectively. A numerical solution of this set of nonlinear equations can be implemented in the form of a homotopy algorithm, which is analogous to [12] and employs the smooth parameter dependence of the stabilizing solutions of DAREs [17] and the vectorization of matrices [11]. The homotopy algorithm for the ﬁltering problem, considered in the present paper, and its convergence will be discussed elsewhere.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	settings based on a single parameter-dependent norm which incorporates the standard H2 and H¥ performance criteria as limiting cases.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	The paper is organized as follows. In Section 2, the mean anisotropy and the anisotropic norm are deﬁned and their basic properties are outlined for convenience. Section 3 formulates the a-anisotropic optimal ﬁltering problem. In Section 4, a sufﬁcient saddle point condition is provided for optimality of an estimator in this problem. Section 5 speciﬁes ﬁnite-dimensional estimators and noise shaping ﬁlters among which the saddle point is being sought. In Section 6, equations are obtained for a worst-case noise shaping ﬁlter against a given ﬁnite-dimensional estimator. Section 7 derives equations for the weighted H2-optimal estimator against a ﬁnite- dimensional shaping ﬁlter, and these results are combined in Section 8 which summarizes a set of algebraic equations for ﬁnding an a-anisotropic optimal estimator.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	coincides with the standard H2-optimal ﬁltering problem in view of the left-most equality in (8). On the other hand, the limit on the right-hand side of (8) suggests that, for large values of the mean anisotropy level a, the problem (14) approaches the H¥ -optimal ﬁltering problem. Note that, irrespective of whether the minimum in the problem (14) is achievable, the quantity \|\|\|a is a nondecreasing concave function of a > 0 as the lower envelope of such infE E ∈ functions.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	denote the set of mean square optimal estimators which minimize the RMS value of the esti- mation errors in (12) for the disturbance W = GV , thereby solving the weighted H2-optimal ﬁltering problem. The following lemma is similar to [25, Lemma 1].
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	Proof. The assertion of the lemma is a corollary from the results of [24, Section 5] (see also (cid:4) [5, Lemmas 5 and 6, Theorem 4]).
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	with a 2n-dimensional internal state (X X). By substituting the covariance relations (53) into , (46)–(48), it follows that the matrix P with necessity satisﬁes the DARE (38)–(40). Moreover, ∗ is an admissible solution the estimator E of the Riccati equation. Hence, in view of the above mentioned uniqueness of such solution, = K(S, L), M = P(S, L), K P = M(S, L) which, under the assumptions of the lemma, implies ∗ ∗ ∗ = M. In order to complete the proof of the lemma, we will need the following = K, M that K ∗ ∗ technical result.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	Now, by recalling (19) and applying Lemma 4 to (54), it follows that X = = EK,M. Therefore, the estimator, described in Lemma 3, indeed satisﬁes (42), which completes (cid:4) the proof.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	The following theorem combines the results of Sections 4–7 and provides a set of equations for ﬁnding an optimal estimator in the a-anisotropic ﬁltering problem.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	Proof. Under the assumptions of the theorem, Lemmas 2 and 3 imply that the estimator EK,M and the shaping ﬁlter GK,S,L satisfy the conditions of Lemma 1, whence the assertion of the (cid:4) theorem follows.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	q=0 = 0, (cid:12) From (34), it follows that, in the case a = 0, the equations of Theorem 1 lead to the the steady- (cid:12) state gain matrices K0, M0 of the standard Kalman ﬁlter.
Processed_Anisotropy-based_optimal_filtering_in_linear_discr.txt	I.G.Vladimirov, and A.P.Kurdjukov, Stochastic approach to H¥ - optimization, Proc. 33rd CDC, Florida, USA, December 14–16, 1994, vol. 3, pp. 2249– 2250.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Weron, 2018). In this paper, we propose a model for probabilistic EPF based on forecast averaging.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	the cyclic daily structure, as well as quadratic terms for both temperature and wind speed.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	with the initial rolling training windows and price classifications, as we will discuss in section 3.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	vertical dashed lines mark the beginning of the rolling windows, which are identified and labelled within the figure.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	quantiles based on a series of explanatory variables, such as time of day and weather conditions.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	to match the longest training length. We will discuss the varying training lengths in section 5.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	we focus on predicting the intra-day variation and mild spiking that occurs with relative frequency.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	using the rolling average price to define the spikes will be influenced by the spikes.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	different rolling windows, capturing the degree to which the price magnitude is a statistical outlier.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	rolling window values, we take the average of the trailing quantiles for the negative spike filtering.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	can be shown to result in predictions of the conditional quantile (Koenker & Hallock, 2001).
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	the GEFcom 2017 competition using some forms of quantile forest (Hong, Xie & Black, 2019).
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	applied to the explanatory variables. Our prediction is the mean of the values in this subspace.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	We can think of this prediction as a weighted sum of the training data yi.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	where ωi(Xt) is 1/s for the s points in the relevant terminal node, 0 elsewhere.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	The ‘forest’ aspect of RF is related to the use of bootstrap aggregation to reduce model variance.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	utilised the AR residual values to increase the robustness of parameter variation in point forecast.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	mined by the AR residual prediction of the post-smoothing 0.5 quantile prediction of that QRF.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	direction of these developments and the main ensemble tool in this study is quantile regression.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	of problems for forecasters, one of which is the selection of the appropriate training period lengths.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	study. It is derived based on the classic CDF approach described in (Lauret et al., 2019).
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	cance in a one-sided test (the median forecasts are identical across these models by construction).
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	together with the averaging routines. Boldface values indicate the top-performing model for a specific interval.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	few hours, its predictions become significantly unstable in the second half of the forecast horizon.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Bordignon, S., Bunn, D., Lisi, F., & Nan, F. (2013). Combining day-ahead forecasts for british electricity prices.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Crane, D. B., & Crotty, J. R. (1967). A two-stage forecasting model: Exponential smoothing and multiple regression.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Dinh, N. T., Pourmousavi, S. A., Karimi-Arpanahi, S., Kumar, Y. P. S., Guo, M., Abbott, D., & Liisberg, J.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Drucker, H., Burges, C. J. C., Kaufman, L., Smola, A., & Vapnik, V. (1996). Support vector regression machines, .
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Koenker, R., & Hallock, K. F. (2001). Quantile regression. Journal of economic perspectives, 15 , 143–156.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Maciejowska, K., & Nowotarski, J. (2016). A hybrid model for gefcom2014 probabilistic electricity price forecasting.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Meinshausen, N., & Ridgeway, G. (2006). Quantile regression forests. Journal of Machine Learning Research, 7 .
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Raviv, E., Bouwman, K. E., & Van Dijk, D. (2015). Forecasting day-ahead electricity prices: Utilizing hourly prices.
Processed_A_probabilistic_forecast_methodology_for_volatile_.txt	Wang, Y., Zhang, N., Tan, Y., Hong, T., Kirschen, D. S., & Kang, C. (2019). Combining probabilistic load forecasts.
Processed_A_Closed-Form_Solution_for_the_Finite_Length_Const.txt	Abstract— In this paper, a closed-form solution minimizing the Godard or Constant Modulus (CM) cost function under the practical conditions of ﬁnite SNR and ﬁnite equalizer length is derived. While previous work has been reported by Zeng et al., IEEE Trans. Information Theory. 1998, to establish the link between the constant modulus and Wiener receivers, we show that under the Gaussian approximation of intersymbol interference at the output of the equalizer, the CM ﬁnite-length receiver is equivalent to the nonblind MMSE equalizer up to a complex gain factor. Some simulation results are provided to support the Gaussian approximation assumption.
Processed_A_Closed-Form_Solution_for_the_Finite_Length_Const.txt	equalizer been proposed in the noisy case. In this paper, a closed-form solution for the ﬁnite-length CM equalizer for a non-minimum phase channel in the presence of additive white Gaussian noise is given. The CM equalizer is shown to be equivalent to the MMSE equalizer up to a complex gain factor. In order to illustrate the theoretical result, a simulation for 4PSK signals is performed, where a comparison of MMSE and CM equalizers, using gradient stochastic algorithms, is implemented.
Processed_A_Closed-Form_Solution_for_the_Finite_Length_Const.txt	Fig. 1. Channel- equalizer block diagram. The cascade channel-equalizer is known as the global response.
Processed_A_Closed-Form_Solution_for_the_Finite_Length_Const.txt	Since MMSE and CM equalizers are identical up to a complex factor, it is theoretically possible to reach the MMSE equalizer performance with a blind receiver composed of a CM equalizer followed by a correction gain and a phase rotator, as depicted in Fig. 2.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	With the prominence of artiﬁcial intelligence, digital content is not only consumed by humans but also by computer pro- grams. This software often analyzes content in different ways, according to their purpose. Depending on their task, only a subset of the information available might be necessary. More- over, the information required can be represented in a more suitable way for the computer program that does not necessar- ily resemble its original natural representation, often required by humans to consume such content.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	which base representations are shared among multiple tasks and only incremental amounts of information are required for more speciﬁc tasks.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Creating learnable tasks that make use of different streams of information in which some are ﬁtted for different purposes is a challenge [3]. The lower-dimensional manifold induced by a particular task might not be readily usable by a differ- ent task. Translating representations from one manifold to another, such that the maximum amount of information is us- able in a secondary task, is limited by the modelling capac- ity of the transformation and the data available [4, 5]. Con- ditional and residual coding have prevailed as two different approaches to incorporate side information in learnable com- pression settings. These approaches can leverage dedicated learnable transformations to explicitly transfer information to the target domain.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	We limit our ﬁndings to a common setting in which we have an image reconstruction task and a computer vision task whose representation is shared with the former. This conﬁg- uration is referred as scalable image coding for humans and machines [3]. We present conditional and residual approaches for scalable learnable compression in which we transform the representations to share a common feature space. We derive baselines for these approaches and empirically compare them. Our experiments perform reconstruction of different datasets using representations for semantic image segmentation and object detection. We also present an entropy model with in- creased modelling potential suitable for conditional coding.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	In learnable compression, an information bottleneck is in- duced on an intermediate representation between the input and the output [6]. Successful approaches follow a variational framework [7] in which a hyper-prior representation learns the dependencies between the different factors of a latent rep- resentation and operates as side information [8, 9, 10, 11]. An auto-regressive entropy model is used to induce the informa- tion bottleneck and to entropy code the learnt representations. Recent work on scalable coding for humans and ma- chines apply the ideas of learnable compression to both the In reconstruction and computer vision tasks [12, 3, 13].
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Fig. 1: Overall architecture of the residual and conditional methods. The dotted line signiﬁes that the enhancement network does not affect the base network. The conditional entropy decoder models H(Yc\|Yt).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	these approaches, the reconstruction task uses a dedicated and a shared representation. These representations are con- catenated after being decoded, and are used as input for a reconstruction model. Through rate-distortion optimization, this approach could create independent representations with no much redundancy of information between them, but the results of [3] and [13] show considerable redundancy. This work focuses primarily on efﬁciently coding the dedicated representation for the reconstruction task. In the conditional approach, the dedicated representation contains all informa- tion relevant to reconstruction but the uncertainty resolved by the shared representation is exploited during coding. In the residual approach, the information of the shared representa- tion is removed from the target representation before coding it, and then added back down the pipeline after decoding.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Residual and conditional coding in the context of learn- able compression has been explored before for video com- pression [14, 15, 16, 17]. Our formulation is different in that the prediction is completely explained by the original signal and as such, the information of the residual cannot increase, whereas in video compression this can occur since the pre- vious frame is used to compute the prediction. In [16], it is shown that learnable conditional coding often requires a transformation of the side information, potentially resulting in information loss to a degree where a residual approach could outperform the conditional approach. In this work, we pro- pose to transform the side information in both approaches and show how this can improve the performance of the residual approach with respect to the conditional approach.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	For an input image X ∈ RCx×H×W , a lossily-compressed base representation Yb = fb(X) is learned as to minimize the distortion Db = EX [db(gb(Yb), T )] with respect to a given computer vision target T , a task distortion function db(·, ·), and a learnable decoding function gb(·).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	In the conditional setting, a lossily-compressed enhance- ment representation Yc = fe(X) is learned to minimize the b distortion Dc = E[de( X = ge(Yc), using an image reconstruction distortion function de(·, ·) and a learnable de- coder ge(·). All information used for the reconstruction task is contained in Yc, and the information contained in Yb is uti- lized to efﬁciently code Yc. Conditional coding effectively models H(Yc\|Yt), where Yt = hc(Yb) is a learnable trans- formation of Yb that intuitively has a feature space similar to that of the enhancement representation Yc, so that their sim- ilarities can be exploited. Any information that reduces the conditional entropy should be maintained in Yt since there is no rate penalty on its rate.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Many of the existing entropy models for learnable com- pression support the conditional coding of the target repre- sentation given the hyper-prior representation [8]. Recent en- tropy models extend these ideas to efﬁciently capture the spa- tial and dimensional dependencies of these representations, by grouping factors together [18], and reorganizing and par- allelizing the decoding order of the spatial locations [19]. In this work, we utilize our conditional information in a similar fashion, but we increase the modelling capacity by augment- ing its receptive ﬁeld and adding scaled residual connections.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	This bound is tight when I(Yc; Yt) = 0, which corresponds to Yc and Yt being independent.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	We provide an upper baseline for the conditional approach by using a standalone enhancement representation Ye gener- ated without relying on any side information, and measuring ˆH(Ye), where ˆH(·) is an entropy estimate. As a lower base- line we use ˆH(Yb)+ ˆH(Ye). This is motivated by considering that Ye can be more efﬁcient as a task representation than Yc, and by the bounds in (1) and (2).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	H(X\|Xp) = H(Xr + Xp\|Xp) = H(Xr\|Xp) = H(Xr) − I(Xp; Xr) ≤ H(Xr) = H(X\|Xp) + I(Xp; Xr).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Here (3) uses the fact that having observed Xp, the only un- certainty in Xr + Xp is due to Xr. The inequality in (4) uses the non-negativity of mutual information. H(Xr) is rewrit- ten in (5) using the deﬁnition of mutual information and once again the fact that given Xp, the only uncertainty in X − Xp is due to X.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	The term I(Xp; Xr) in (4) acts as a penalty term on the residual formulation. To minimize it, the residual Xr and the prediction Xp must be as independent from each other as pos- sible. This can be achieved when Xp collapses values in X so that H(Xr) decreases, or when Xp produces a constant value for different values in X, reducing H(Xp). Reducing H(Xp) increases H(X\|Xp), which in turn could have the adverse ef- fect of increasing H(Xr), as shown in (5).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	In our proposed method we train to minimize Dr and ˆH(Yr). By extension, we also minimize ˆH(Xr). As shown in (5), this reduces both H(X\|Xp) and I(Xp; Xr). Hence, this optimization procedure encourages the learnable function hr(·) to create a representation Xp that recovers the input X as accurately as possible, while at the same time being as in- dependent as possible from the resulting residual Xr.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Note that enforcing the similarity of Xp and the original input X may not be an optimal procedure, since even though such an optimization will decrease H(X\|Xp), it may lead to an increase in I(Xp; Xr). This explains why in preliminary experiments, we found that having a function hr(·) that ex- plicitly reconstructs X does not perform as well as our pro- posed method.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Due to the previous considerations stemming from our proposed method, we ﬁnd that H(Yc\|Yt) = H(Yr) can be easier to achieve. This motivates us to compare ˆH(Yb) + ˆH(Yr) against the same baselines used in the conditional ap- proach.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	To conditionally code a representation Yc that exploits as much information as possible from Yt, we model the spatial and dimensional dependencies between and within repre- sentations using a CNN [18]. Our proposed entropy model strikes a balance between complexity and accuracy. We group channels with a ﬁxed size K [18]. Within each group, the same location across channels are processed in parallel, using as context all locations in the previous groups and all previous locations across all channels of the current group, within the receptive ﬁeld of the convolutional layer. Similarly to [8], lo- cations in the spatial domain are processed in a top-to-bottom, left-to-right fashion. The Markov property is enforced by a mask applied to the convolution kernels. Fig. 2a shows the kernel mask for a single output channel of a layer.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Unlike previous work, the CNN architecture of our en- tropy model has scalable residual connections and deeper lay- ers with kernels sizes larger than one for its auto-regressive convolutions. This removes some of the modelling limita- tions imposed by similar entropy models. The CNN archi- tecture has blocks of three layers in which the input channels are scaled up, transformed in a higher-dimensional space, and scaled down back to the original number of channels. The residual connections are introduced in-between these blocks, such that the inputs can be re-scaled differently across the channel dimension. To maintain the Markov property when the number of channels changes, the group sizes are re-scaled accordingly and the channels can only change in multiples of M . Fig. 2b shows the architecture overview of a single block in the CNN.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	The conditional representation is available as another channel group and is transformed by the CNN in the same way as other groups, except that its context is restricted to the receptive ﬁeld within that group. All elements of the conditioned representation have access to this group.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Similarly to [8] and more recent works, the predictions of the entropy model correspond to the means W and scales Σ of a univariate Gaussian distribution assigned to each ele- ment in the representation. The symbols are obtained by Q = ⌊Y −W ⌉, and the corresponding probability is PN (0,Σ)[\|Q\| ≤ \|Q\| ± 1/2]. During training, the rounding operation is simu- lated by adding uniform noise U(−1/2, 1/2).
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Fig. 2: Entropy model overview. (a) The convolution has kernel size 3 × 3 and the input is 12 × 4 × 4. The conditional input has size 3 × 4 × 4 and there are K = 4 groups. With an input padding and a stride of 1, this is the 7th step of the convolution.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	components from the coder, introduced as a hyper-prior in [8]. We also remove the attention layers introduced in [10]. To reduce the memory footprint and speed up the training proce- dure, we incrementally reduce the channels in the ﬁrst layers of the analyzers and the last layers of the synthesizers.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	for the conditional and residual approaches, respectively. During training, either the base network remains frozen or the gradients from the reconstruction network do not ﬂow into the base network.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	The base and enhancement tasks have this same architec- ture. The base generates a representation with the same di- mensionality and resolution as the input X but reconstruction of the input is not enforced. This representation is the input for the computer vision model. The coder and the computer vision model are trained together end-to-end.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	In the conditional approach, hc(·) is a traditional CNN composed of blocks with residual connections between them. Each block has three convolutional layers that perform a transformation in a higher dimensionality and scales the out- put back to a lower dimensionality. The ﬁrst half of the blocks maintain the same dimensionality as Yb, while the second half transitions to the same dimensionality as Yc, ob- taining Yt. The resolution is maintained across the network as both Yb and Yc have the same resolution. In the residual approach, hr(·) uses the same architecture as our synthetiz- ers, to transform Yb into Xp. The synthetizer upscales the representation to match the resolution and dimensionality of X.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	We conduct experiments to analyze the rate-distortion per- formance of the proposed conditional and residual methods for scalable coding and compare them against the proposed baselines. We perform two sets of experiments: one using semantic segmentation as the computer vision task on the Cityscapes dataset [20], and another using object detection as the computer vision task on the COCO 2017 dataset [21]. We ﬁrst train the base representation on the computer vi- sion task to obtain fb(·) and gb(·) for rate-distortion points under different values of λb. We choose a model correspond- ing to a point on the rate-distortion curve that achieves rea- sonable distortion and subsequently use it to generate the rep- resentations Yb for the conditional and residual approaches. The upper baseline is created by training the reconstruction task with no side information. We use the same architecture for the analysis and synthesis transforms as the other models. The entropy model for the upper baseline has the same archi- tecture as the one used in the residual approach. The lower baseline is obtained by adding the rate of the base representa- tion used for the conditional and residual approaches.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Fig. 3: Scalable coding results. The purple lines represent the performance attained with λb = 0 and β = 0.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	under low-compression settings and use its weights as initial- ization for the models trained to obtain the rest of the curve. The parameters are updated using Adam at a learning rate of 10−4. We train models with early stopping but ﬁrst decay the learning rate by a factor of 0.75 if a plateau is reached.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Cityscapes is a set of images of urban scenes for semantic un- derstanding [20]. We use DeepLabV3 [22] as the computer vision model for segementation with MobileNetV3 [23] as a back-end. Here, db(·, ·) is the per-pixel multi-class cross- entropy, although we report the mean intersection over union (mIoU) metric. We set β = 0.1 and report the results on the validation dataset. For data augmentation, we use ran- dom crops of 768 × 768 pixels, random horizontal ﬂips and color jittering. The front-end of the model corresponding to the coder is trained with Adam using a learning rate of 10−4, while the classiﬁer is trained with stochastic gradient descent using momentum and a learning rate of 10−2. A ℓ2 loss is added to the weights of the classiﬁer to prevent over-ﬁtting, with a scale factor of 10−4.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Fig. 3a shows the rate-distortion curves for the condi- tional and residual approach. We notice that these lines lie in between the baselines, producing rate-distortion points that respect them. Compared to the rate-distortion curve of the lower baseline, the conditional approach has a BD-Rate [24] of −16.56%, whereas the residual approach achieves a −14.6% rate reduction. Thus, the conditional approach per- forms marginally better than the residual approach in terms of BD-Rate. Looking at the ratio between these BD-Rate scores and the BD-Rate score achieved by the upper baseline, we can compute the percentage of the base representation utilized. As such, the conditional approach uses 43.01% of the side information rate, whereas the residual approach uses 37.91%. In the lowest-compression settings under both approaches, the utilization is higher.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	Fig. 3c shows the rate-distortion performance of the base task. The chosen β value places a penalty on both the rate and the task performance but allows the base representation to be exploited by the architecture. We attribute the small imperfections in this rate-distortion curve to the choice of β and the limitations of the training algorithm.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	COCO 2017 has 123,287 domain-agnostic images for object detection and segmentation [21]. We use Faster R-CNN [25] for object detection with ResNet-50 [26] as the back-end. For this task, db(·, ·) is the sum of the different loss functions employed by this architecture. We report the mean average precision (mAP) metric computed according to [21], and set β = 0.05. As data preprocessing for training, we use random horizontal ﬂips and generate batches with similar aspect ratios grouped in 3 clusters. Images inside a batch are resized to the minimum size and the bounding boxes are adjusted accord- ingly when training the computer vision task. When training for reconstruction, the images in a batch are center-cropped to their minimum size. All weights are trained with Adam using a learning rate of 10−4.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	As shown in Fig. 3b, the performance of both approaches is comparable, with a −4.14% and a −2.47% BD-Rate im- provement over the lower baseline, for the conditional and residual methods, respectively. We achieve a utilization of the base in terms of BD-Rate of 49.24% and 29.32% for the con- ditional and residual approaches, respectively. Fig. 3d shows the base-distortion performance of the base task. Compared to semantic segmentation on Cityscapes, the task model is better at reaching the uncompressed task performance. Also, for a similar distortion penalty, this task uses more rate. The rate-distortion curves obtained by the reconstruction task on the COCO dataset are almost an order of magnitude larger than the ones from Cityscapes. This can be explained by the simplicity of the content in the images found in Cityscapes, and the higher amount of artifacts found in the COCO dataset due to compression.
Processed_Conditional_and_Residual_Methods_in_Scalable_Codin.txt	We present conditional and residual methods for scalable cod- ing for humans and machines. Our experiments show that the proposed architectures for conditional and residual coding perform similarly and that the rate-distortion performance is within the presented baselines or operational bounds. In addi- tion, the proposed conditional entropy model is able to match the performance of the residual method.
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	Abstract. Three-dimensional (3D) magnetic nulls are the points where magnetic field vanishes and are preferential sites for magnetic reconnection: a process which converts magnetic energy into heat and accelerates charged particles along with a rearrangement of magnetic field lines. In the solar corona, the reconnections manifest as coronal transients including solar flares, coronal mass ejections and coronal jets. The nulls are generally found to be collocated with complex active regions on the solar photosphere. Extrapolation of magnetic field from corresponding photospheric magnetogram indicate an abundance of these nulls in the solar atmosphere. Nevertheless, their generation is still not well understood. Recently, Maurya et al. (2023) have demonstrated magnetic reconnection to be a cause for generation and annihilation of magnetic nulls through magnetohydrodynamics simulation, where the initial magnetic field is idealized to have a single proper radial null. This article further extends the study in a more realistic scenario where the initial magnetic field is constructed by extrapolating photospheric magnetogram data and hence, incorporates field line complexities inherent to a complex active region. For the purpose, the active region NOAA 11977 hosting a C6.6 class flare is selected. The simulation is initiated using non-force-free extrapolated magnetic field from the photospheric vector magnetogram at around 02 : 48 : 00 UT on 17 February 2014, 16 minutes before the flare peak. The generation, annihilation and dynamics of nulls are explored by a complimentary usage of trilinear null detection technique and tracing of magnetic field line dynamics. It is found that the nulls can spontaneously generate/annihilate in pairs while preserving the topological degree and can have observational implications like footpoint brightenings. Magnetic reconnection is found to be the cause of such generation and annihilation.
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	Three-dimensional (3D) nulls are locations in space where the magnetic field is zero. Contemporary researches in theory, observation and numerical simulation have shown the importance of 3D magnetic nulls in natural phenomena involving magnetic reconnection: a process where magnetic field line connectivity changes along with generation of heat and acceleration of charged particles. 3D nulls are preferential sites for magnetic reconnections (Priest and Forbes, 2000; Birn and Priest, 2007; Pontin et al., 2007; Zweibel and Yamada, 2009; Yamada et al., 2010; Pontin and Priest, 2022) which, in turn, are responsible for various transient events in the solar corona—including persistent null point brightening (Cheng et al., 2023), solar flares (Aulanier et al., 2000; Fletcher et al., 2001; Masson et al., 2009), coronal jets (Wyper et al., 2018; Nayak et al., 2019) and coronal mass ejections (Lynch et al., 2008; Longcope, 1996).
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	and remains conserved (Hornig and Schindler, 1996; Pontin and Priest, 2022) in any evolution. Any credible simulation targeted to explore generation or annihilation of nulls must satisfy this stringent conservation. Additionally, nulls that enter or exit the computational domain needs to be accounted for as they may seemingly violate the conservation of topological degree.
Processed_Generation_and_annihilation_of_three_dimensional_m.txt	point. With the onset of reconnection at the null, magnetic field lines are transferred across the separatrices from one magnetic domain to another (Pontin et al., 2013).

Subsets and Splits