Breaking The Curse of Dimensionality With Convex Neural Networks

Journal of Machine Learning Research 18 (2017) 1-53 Submitted 12/14; Revised 10/16; Published 4/17
Breaking the Curse of Dimensionality

with Convex Neural Networks
Francis Bach [email protected]

INRIA - Sierra Project-team
Département d’Informatique de l’Ecole Normale Supérieure (UMR CNRS/ENS/INRIA)
2, rue Simone Iff
75012 Paris, France
Editor: Yoshua Bengio
Abstract
We consider neural networks with a single hidden layer and non-decreasing positively ho-
mogeneous activation functions like the rectified linear units. By letting the number of
hidden units grow unbounded and using classical non-Euclidean regularization tools on
the output weights, they lead to a convex optimization problem and we provide a de-
tailed theoretical analysis of their generalization performance, with a study of both the
approximation and the estimation errors. We show in particular that they are adaptive
to unknown underlying linear structures, such as the dependence on the projection of the
input variables onto a low-dimensional subspace. Moreover, when using sparsity-inducing
norms on the input weights, we show that high-dimensional non-linear variable selection
may be achieved, without any strong assumption regarding the data and with a total num-
ber of variables potentially exponential in the number of observations. However, solving
this convex optimization problem in infinite dimensions is only possible if the non-convex
subproblem of addition of a new unit can be solved efficiently. We provide a simple geo-
metric interpretation for our choice of activation functions and describe simple conditions
for convex relaxations of the finite-dimensional non-convex subproblem to achieve the same
generalization error bounds, even when constant-factor approximations cannot be found.
We were not able to find strong enough convex relaxations to obtain provably polynomial-
time algorithms and leave open the existence or non-existence of such tractable algorithms
with non-exponential sample complexities.
Keywords: Neural networks, non-parametric estimation, convex optimization, convex
relaxation.
1. Introduction
Supervised learning methods come in a variety of ways. They are typically based on local
averaging methods, such as k-nearest neighbors, decision trees, or random forests, or on
optimization of the empirical risk over a certain function class, such as least-squares re-
gression, logistic regression or support vector machine, with positive definite kernels, with
model selection, structured sparsity-inducing regularization, or boosting (see, e.g., Györfi
and Krzyzak, 2002; Hastie et al., 2009; Shalev-Shwartz and Ben-David, 2014, and references
therein).
c
2017 Francis Bach.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided
at http://jmlr.org/papers/v18/14-546.html.
Bach
Most methods assume either explicitly or implicitly a certain class of models to learn
from. In the non-parametric setting, the learning algorithms may adapt the complexity
of the models as the number of observations increases: the sample complexity (i.e., the
number of observations) to adapt to any particular problem is typically large. For example,
when learning Lipschitz-continuous functions in Rd , at least n = Ω(ε− max{d,2} ) samples are
needed to learn a function with excess risk ε (von Luxburg and Bousquet, 2004, Theorem
15). The exponential dependence on the dimension d is often referred to as the curse of
dimensionality: without any restrictions, exponentially many observations are needed to
obtain optimal generalization performances.
At the other end of the spectrum, parametric methods such as linear supervised learning
make strong assumptions regarding the problem and generalization bounds based on esti-
mation errors typically assume that the model is well-specified, and the sample complexity
to attain an excess risk of ε grows as n = Ω(d/ε2 ), for linear functions in d dimensions
and Lipschitz-continuous loss functions (Shalev-Shwartz and Ben-David, 2014, Chapter 9).
While the sample complexity is much lower, when the assumptions are not met, the methods
underfit and more complex models would provide better generalization performances.
Between these two extremes, there are a variety of models with structural assumptions
that are often used in practice. For input data in x ∈ Rd , prediction functions f : Rd → R
may for example be parameterized as:
(a) Affine functions: f (x) = w> x + b, leading to potential severe underfitting, but easy
optimization and good (i.e., non exponential) sample complexity.
Pd
(b) Generalized additive models: f (x) = j=1 fj (xj ), which are generalizations of the
above by summing functions fj : R → R which may not be affine (Hastie and Tibshi-
rani, 1990; Ravikumar et al., 2008; Bach, 2008a). This leads to less strong underfitting
but cannot model interactions between variables, while the estimation may be done
with similar tools than for affine functions (e.g., convex optimization for convex losses).
P
(c) Nonparametric ANOVA models: f (x) = A∈A fA (xA ) for a set A of subsets of
{1, . . . , d}, and non-linear functions fA : RA → R. The set A may be either given (Gu,
2013) or learned from data (Lin and Zhang, 2006; Bach, 2008b). Multi-way interac-
tions are explicitly included but a key algorithmic problem is to explore the 2d − 1
non-trivial potential subsets.
(d) Single hidden-layer neural networks: f (x) = kj=1 σ(wj> x+bj ), where k is the number
P
of units in the hidden layer (see, e.g., Rumelhart et al., 1986; Haykin, 1994). The
activation function σ is here assumed to be fixed. While the learning problem may
be cast as a (sub)differentiable optimization problem, techniques based on gradient
descent may not find the global optimum. If the number of hidden units is fixed, this
is a parametric problem.
(e) Projection pursuit (Friedman and Stuetzle, 1981): f (x) = kj=1 fj (wj> x) where k is
P
the number of projections. This model combines both (b) and (d); the only difference
with neural networks is that the non-linear functions fj : R → R are learned from data.
The optimization is often done sequentially and is harder than for neural networks.
2
Breaking the Curse of Dimensionality with Convex Neural Networks
(e) Dependence on a unknown k-dimensional subspace: f (x) = g(W > x) with W ∈ Rd×k ,
where g is a non-linear function. A variety of algorithms exist for this problem (Li,
1991; Fukumizu et al., 2004; Dalalyan et al., 2008). Note that when the columns of W
are assumed to be composed of a single non-zero element, this corresponds to variable
selection (with at most k selected variables).
In this paper, our main aim is to answer the following question: Is there a single
learning method that can deal efficiently with all situations above with prov-
able adaptivity ? We consider single-hidden-layer neural networks, with non-decreasing
homogeneous activation functions such as
σ(u) = max{u, 0}α = (u)α+ ,
for α ∈ {0, 1, . . .}, with a particular focus on α = 0 (with the convention that 00 = 0), that
is σ(u) = 1u>0 (a threshold at zero), and α = 1, that is, σ(u) = max{u, 0} = (u)+ , the
so-called rectified linear unit (Nair and Hinton, 2010; Krizhevsky et al., 2012). We follow
the convexification approach of Bengio et al. (2006); Rosset et al. (2007), who consider
potentially infinitely many units and let a sparsity-inducing norm choose the number of
units automatically. This leads naturally to incremental algorithms such as forward greedy
selection approaches, which have a long history for single-hidden-layer neural networks (see,
e.g., Breiman, 1993; Lee et al., 1996).
We make the following contributions:
– We provide in Section 2 a review of functional analysis tools used for learning from
continuously infinitely many basis functions, by studying carefully the similarities and
differences between L1 - and L2 -penalties on the output weights. For L2 -penalties,
this corresponds to a positive definite kernel and may be interpreted through random
sampling of hidden weights. We also review incremental algorithms (i.e., forward
greedy approaches) to learn from these infinite sets of basis functions when using
L1 -penalties.
– The results are specialized in Section 3 to neural networks with a single hidden layer
and activation functions which are positively homogeneous (such as the rectified linear
unit). In particular, in Sections 3.2, 3.3 and 3.4, we provide simple geometric interpre-
tations to the non-convex problems of additions of new units, in terms of separating
hyperplanes or Hausdorff distance between convex sets. They constitute the core po-
tentially hard computational tasks in our framework of learning from continuously
many basis functions.
– In Section 4, we provide a detailed theoretical analysis of the approximation prop-

erties of (single hidden layer) convex neural networks with monotonic homogeneous
activation functions, with explicit bounds. We relate these new results to the exten-
sive literature on approximation properties of neural networks (see, e.g., Pinkus, 1999,
and references therein) in Section 4.7, and show that these neural networks are indeed
adaptive to linear structures, by replacing the exponential dependence in dimension
by an exponential dependence in the dimension of the subspace of the data can be
projected to for good predictions.
3
Bach

Functional form Generalization bound

−1/(d+3)
No assumption n log n

w> x d1/2 · n−1/2

Affine function
+ b

Pk > wj ∈ R d kd1/2 · n−1/4 log n

Generalized additive model j=1 fj (wj x),

Pk > kd1/2 · n−1/2

Single-layer neural network j=1 ηj (wj x + bj )+

Pk > wj ∈ R d kd1/2 · n−1/4 log n

Projection pursuit j=1 fj (wj x),

f (W > x) , W ∈ Rd×s d1/2 · n−1/(s+3) log n

Dependence on subspace

Table 1: Summary of generalization bounds for various models. The bound represents the
expected excess risk over the best predictor in the given class. When no assumption
is made, the dependence in n goes to zero with an exponent proportional to 1/d
(which leads to sample complexity exponential in d), while making assumptions
removes the dependence of d in the exponent.
– In Section 5, we study the generalization properties under a standard supervised learn-

ing set-up, and show that these convex neural networks are adaptive to all situations
mentioned earlier. These are summarized in Table 1 and constitute the main statis-
tical results of this paper. When using an `1 -norm on the input weights, we show in
Section 5.3 that high-dimensional non-linear variable selection may be achieved, that
is, the number of input variables may be much larger than the number of observa-
tions, without any strong assumption regarding the data (note that we do not present
a polynomial-time algorithm to achieve this).
– We provide in Section 5.5 simple conditions for convex relaxations to achieve the
same generalization error bounds, even when constant-factor approximation cannot
be found (e.g., because it is NP-hard such as for the threshold activation function and
the rectified linear unit). We present in Section 6 convex relaxations based on semi-
definite programming, but we were not able to find strong enough convex relaxations
(they provide only a provable sample complexity with a polynomial time algorithm
which is exponential in the dimension d) and leave open the existence or non-existence
of polynomial-time algorithms that preserve the non-exponential sample complexity.
2. Learning from Continuously Infinitely Many Basis Functions

In this section we present the functional analysis framework underpinning the methods
presented in this paper, which learn for a potential continuum of features. While the
formulation from Sections 2.1 and 2.2 originates from the early work on the approxima-
4
tion properties of neural networks (Barron, 1993; Kurkova and Sanguineti, 2001; Mhaskar,
2004), the algorithmic parts that we present in Section 2.5 have been studied in a variety
of contexts, such as “convex neural networks” (Bengio et al., 2006), or `1 -norm with infi-
nite dimensional feature spaces (Rosset et al., 2007), with links with conditional gradient
algorithms (Dunn and Harshbarger, 1978; Jaggi, 2013) and boosting (Rosset et al., 2004).
In the following sections, note that there will be two different notions of infinity: in-
finitely many inputs x and infinitely many basis functions x 7→ ϕv (x). Moreover, two
orthogonal notions of Lipschitz-continuity will be tackled in this paper: the one of the pre-
diction functions f , and the one of the loss ` used to measure the fit of these prediction
functions.
2.1 Variation Norm

We consider an arbitrary measurable input space X (this will a sphere in Rd+1 starting
from Section 3), with a set of basis functions (a.k.a. neurons or units) ϕv : X → R, which
are parameterized by v ∈ V, where V is a compact topological space (typically a sphere
for a certain norm on Rd starting from Section 3). We assume that for any given x ∈ X ,
the functions v 7→ ϕv (x) are continuous. These functions will be the hidden neurons in
a single-hidden-layer neural network, and thus V will be (d + 1)-dimensional for inputs of
dimension d (to represent any affine function). Throughout Section 2, these features will
be left unspecified as most of the tools apply more generally.
In order to define our space of functions from X → R, we need real-valued Radon
measures, which are continuous linear forms on the space of continuous functions from V
to R, equipped with the uniform norm (Rudin, 1987; Evans and Gariepy, 1991). For a
Rcontinuous function g : V → R and a Radon measure µ, we will use the standard notation
V g(v)dµ(v) to denote the action of the measure µ on the continuous function g. The norm
of µ is usually referred to as its total variation (such finite total variation corresponds to
having a continuous linear form on the space R of continuous functions), and we denote it
as |µ|(V), and is equal to the supremum of V g(v)dµ(v) over all continuous functions with
values in [−1, 1]. As seen below, when µ has a density with respect to a probability measure,
this is the L1 -norm of the density.
We consider the space F1 of functions f that can be written as
Z
f (x) = ϕv (x)dµ(v),
V
where µ is a signed Radon measure on V with finite total variation |µ|(V).

When V is finite, this corresponds to
X
f (x) = µv ϕv (x),
v∈V
P
with total variation v∈V |µv |, where the proper formalization for infinite sets V is done
through measure theory. R
The infimum of |µ|(V) over all decompositions of f as f = V ϕv dµ(v), turns out to
be a norm γ1 on F1 , often called the variation norm of f with respect to the set of basis
functions (see, e.g., Kurkova and Sanguineti, 2001; Mhaskar, 2004).
5
Bach
Given our assumptions regarding the compactness of V, for any f ∈ F1 , the infimum
defining γ1 (f ) is in fact attained by a signed measure µ, as a consequence of the compactness
of measures for the weak topology (see Evans and Gariepy, 1991, Section 1.9).
In the definition above, if we assume that the signed measure µ has a density with
respect to a fixed probability measure τ with full support on V, that is, dµ(v) = p(v)dτ (v),
then, the variation norm γ1 (f ) is also equal to the infimal value of
Z
|µ|(V) = |p(v)|dτ (v),
V
R
over all integrable functions p such that f (x) = V p(v)ϕv (x)dτ (v). Note however that not
all measures have densities, and that the two infimums are the same as all Radon measures
are limits of measures with densities. Moreover, the infimum in the definition above is not
attained in general (for example when the optimal measure is singular with respect to dτ );
however, it often provides a more intuitive definition of the variation norm, and leads to
easier comparisons with Hilbert spaces in Section 2.3.
Finite number of neurons. If f : X → R is decomposableP into k basis functions, that
is, f (x) = kj=1 ηj ϕvj (x), then this corresponds to µ = kj=1 ηj δ(v = vj ), and the total
P
variation of µ is equal to the `1 -norm kηk1 of η. Thus the function f has variation norm
less than kηk1 or equal. This is to be contrasted with the number of basis functions, which
is the `0 -pseudo-norm of η.
2.2 Representation from Finitely Many Functions

When minimizing any functional J that depends only on the function values taken at a
subset X̂ of values in X , over the ball {f ∈ F1 , γ1 (f ) 6 δ}, then we have a “representer
theorem” similar to the reproducing kernel Hilbert space situation, but also with significant
differences, which we now present.
The problem is indeed simply equivalent to minimizing a functional on functions re-
stricted to X , that is, to minimizing J(f|X̂ ) over f|X̂ ∈ RX̂ , such that γ1|X̂ (f|X̂ ) 6 δ, where
Z
γ1|X̂ (f|X̂ ) = inf |µ|(V) such that ∀x ∈ X̂ , f|X̂ (x) = ϕv (x)dµ(v);
µ V
we can then build a function defined over all X , through the optimal measure µ above.
Moreover, by Carathéodory’s theorem for cones (Rockafellar, 1997), if X̂ is composed
of only n elements (e.g., n is the number of observations in machine learning), the optimal
function f|X̂ above (and hence f ) may be decomposed into at most n functions ϕv , that is,
µ is supported by at most n points in V, among a potential continuum of possibilities.
Note however that the identity of these n functions is not known in advance, and thus
there is a significant difference with the representer theorem for positive definite kernels and
Hilbert spaces (see, e.g., Shawe-Taylor and Cristianini, 2004), where the set of n functions
are known from the knowledge of the points x ∈ X̂ (i.e., kernel functions evaluated at x).
2.3 Corresponding Reproducing Kernel Hilbert Space (RKHS)

We have seen above that if the real-valued measures µ are restricted to have density p with
respect to a fixed probability measure τ with full support on V, that is, dµ(v) = p(v)dτ (v),
6
R
then, the norm γ1 (f ) isRthe infimum of the total variation |µ|(V) = V |p(v)|dτ (v), over all
decompositions f (x) = V p(v)ϕv (x)dτR(v).
We may also define the infimum of V |p(v)|2 dτ (v) over the same decompositions (squared
L2 -norm instead of L1 -norm). It turns out that it defines a squared norm γ22 and that
the function space F2 of functions with finite norm happens to be a reproducing ker-
nel Hilbert space (RKHS). When V is finite, then it is well-known (see, e.g., Berlinet
2 over all vectors µ
P
and Thomas-Agnan,
P 2004, Section 4.1) that the infimum of v∈V vµ
such thatP f = v∈V µv ϕv defines a squared RKHS norm with positive definite kernel
k(x, y) = v∈V ϕv (x)ϕv (y).
We show in Appendix A that for any compact Z set V, we have defined a squared RKHS
norm γ22 with positive definite kernel k(x, y) = ϕv (x)ϕv (y)dτ (v).
V
Random sampling. Note that such kernels are well-adapted to approximations by sam-
pling several basis functions ϕv sampled from the probability measure τ (Neal, 1995; Rahimi
and Recht, 2007). Indeed,Pif we consider m i.i.d. samples v1 , . . . , vm , we may define the ap-
1 m
proximation k̂(x, y) = m i=1 ϕvi (x)ϕvi (y), which corresponds to an explicit feature repre-
sentation. In other words, this corresponds to sampling units vi , using prediction functions
1 Pm
of the form m η ϕ
i=1 i vi (x) and then penalizing by the `2 -norm of η.
When m tends to infinity, then k̂(x, y) tends to k(x, y) and random sampling provides a
way to work efficiently with explicit m-dimensional feature spaces. See Rahimi and Recht
(2007) for a analysis of the number of units needed for an approximation with error ε, typ-
ically of order 1/ε2 . See also Bach (2017) for improved results with a better dependence on
ε when making extra assumptions on the eigenvalues of the associated covariance operator.
Relationship between F1 and F2 . The corresponding RKHS norm is always greater
than the variation norm (because of Jensen’s inequality), and thus the RKHS F2 is included
in F1 . However, as shown in this paper, the two spaces F1 and F2 have very different
properties; e.g., γ2 may be computed easily in several cases, while γ1 does not; also, learning
with F2 may either be done by random sampling of sufficiently many weights or using kernel
methods, while F1 requires dedicated convex optimization algorithms with potentially non-
polynomial-time steps (see Section 2.5).
Moreover, for any v ∈ V, ϕv ∈ F1 with a norm γ1 (ϕv ) 6 1, while in general ϕv ∈ / F2 .
This is a simple illustration of the fact that F2 is too small and thus will lead to a lack
of adaptivity that will be further studied in Section 5.4 for neural networks with certain
activation functions.
2.4 Supervised Machine Learning

Given some distribution over the pairs (x, y) ∈ X ×Y, a loss function ` : Y ×R → R, our aim
is to find a function f : X → R such that the functional J(f ) = E[`(y, f (x))] is small, given
some i.i.d. observations (xi , yi ), i = 1, . . . , n. We consider the empirical risk minimization
framework over a space of functions F, equipped with a norm γ (in our situation, F1 and
F2 , equipped with γ1 or γ2 ). The empirical risk J(f ˆ ) = 1 Pn `(yi , f (xi )), is minimized
n i=1
either (a) by constraining f to be in the ball F δ = {f ∈ F, γ(f ) 6 δ} or (b) regularizing
the empirical risk by λγ(f ). Since this paper has a more theoretical nature, we focus on
constraining, noting that in practice, penalizing is often more robust (see, e.g., Harchaoui
7
Bach
f¯t = arg min hJ ′ (ft ), f i

γ1 (f ) ≤ δ γ1 (f )≤δ
−J ′ (ft ) ft+1
ft
Figure 1: Conditional gradient algorithm for minimizing a smooth functional J on F1δ =

{f ∈ F1 , γ1 (f ) 6 δ}: going from ft to ft+1 ; see text for details.
et al., 2013) and leaving its analysis in terms of learning rates for future work. Since the
functional Jˆ depends only on function values taken at finitely many points, the results
from Section 2.2 apply and we expect the solution f to be spanned by only n functions
ϕv1 , . . . , ϕvn (but we ignore in advance which ones among all ϕv , v ∈ V, and the algorithms
in Section 2.5 will provide approximate such representations with potentially less or more
than n functions).
Approximation error vs. estimation error. We consider an ε-approximate minimizer

1 Pn
of J(f ) = n i=1 `(yi , f (xi )) on the convex set F δ , that is a certain fˆ ∈ F δ such that
ˆ
ˆ fˆ) 6 ε+inf f ∈F δ J(f
J( ˆ ). We thus have, using standard arguments (see, e.g., Shalev-Shwartz
and Ben-David, 2014):

J(fˆ) − inf J(f ) 6 ˆ ) − J(f )| + ε,
inf J(f ) − inf J(f ) + 2 sup |J(f
f ∈F f ∈F δ f ∈F f ∈F δ
that is, the excess risk J(fˆ) − inf f ∈F J(f ) is upper-bounded by a sum of an approximation
ˆ ) − J(f )| and an opti-
error inf f ∈F δ J(f ) − inf f ∈F J(f ), an estimation error 2 supf ∈F δ |J(f
mization error ε (see also Bottou and Bousquet, 2008). In this paper, we will deal with all
three errors, starting from the optimization error which we now consider for the space F1
and its variation norm.
2.5 Incremental Conditional Gradient Algorithms

In this section, we review algorithms to minimize a smooth functional J : L2 (dρ) → R,
where ρ is a probability measure on X . This may typically be the expected risk or the
empirical risk above. When minimizing J(f ) with respect to f ∈ F1 such that γ1 (f ) 6 δ, we
need algorithms that can efficiently optimize a convex function over an infinite-dimensional
space of functions. Conditional gradient algorithms allow to incrementally build a set of
elements of F1δ = {f ∈ F1 , γ1 (f ) 6 δ}; see, e.g., Frank and Wolfe (1956); Dem’yanov and
Rubinov (1967); Dudik et al. (2012); Harchaoui et al. (2013); Jaggi (2013); Bach (2015).
8
Conditional gradient algorithm. We assume the functional J is convex and L-smooth,

that is for all h ∈ L2 (dρ), there exists a gradient J 0 (h) ∈ L2 (dρ) such that for all f ∈ L2 (dρ),
L
0 6 J(f ) − J(h) − hf − h, J 0 (h)iL2 (dρ) 6
kf − hk2L2 (dρ) .
2
When X is finite, this corresponds to the regular notion of smoothness from convex opti-
mization (Nesterov, 2004).
The conditional gradient algorithm (a.k.a. Frank-Wolfe algorithm) is an iterative algo-
rithm, starting from any function f0 ∈ F1δ and with the following recursion, for t > 0:
f¯t ∈ arg min hf, J 0 (ft )iL2 (dρ)
f ∈F1δ
ft+1 = (1 − ρt )ft + ρt f¯t .

2
See an illustration in Figure 1. We may choose either ρt = t+1 or perform a line search
for ρt ∈ [0, 1]. For all of these strategies, the t-th iterate is a convex combination of the
functions f¯0 , . . . , f¯t−1 , and is thus an element of F1δ . It is known that for these two strategies
for ρt , we have the following convergence rate (see, e.g. Jaggi, 2013):
2L
J(ft ) − inf J(f ) 6 sup kf − gk2L2 (dρ) .
f ∈F1δ t + 1 f,g∈F δ
1
When, r2 = supv∈V kϕv k2L2 (dρ) is finite, we have kf k2L2 (dρ) 6 r2 γ1 (f )2 and thus we get a
2 2
convergence rate of 2Lr t+1 .
δ
Moreover, the basic Frank-Wolfe (FW) algorithm may be extended to handle the reg-
ularized problem as well (Harchaoui et al., 2013; Bach, 2013; Zhang et al., 2012), with
similar convergence rates in O(1/t). Also, the second step in the algorithm, where the
function ft+1 is built in the segment between ft and the newly found extreme function,
may be replaced by the optimization of J over the convex hull of all functions f¯0 , . . . , f¯t ,
a variant which is often referred to as fully corrective. Moreover, in our context where V
is a space where local search techniques may be considered, there is also the possibility
of “fine-tuning” the vectors v as wellP(Bengio et al., 2006), that is, we may optimize the
function (v1 , . . . , vt , α1 , . . . , αt ) 7→ J( ti=1 αi ϕvi ), through local search techniques, starting
from the weights (αi ) and points (vi ) obtained from the conditional gradient algorithm.
Adding a new basis function. The conditional gradient algorithm presented above
relies on solving at each iteration the “Frank-Wolfe step”:
max hf, giL2 (dρ) .
γ(f )6δ
for gR = −J 0 (ft ) ∈ L2 (dρ). For the norm γ1 defined through an L1 -norm, we have for
f = V ϕv dµ(v) such that γ1 (f ) = |µ|(V):
Z Z Z
hf, giL2 (dρ) = f (x)g(x)dρ(x) = ϕv (x)dµ(v) g(x)dρ(x)
X X V
Z Z
= ϕv (x)g(x)dρ(x) dµ(v)
V X
Z

6 γ1 (f ) · max ϕv (x)g(x)dρ(x),
v∈V X
9
Bach
with equality if and only if µ = µ+ − µ− with µ+ and µR− two non-negative measures, with
µ+ (resp. µ− ) supported in the set of maximizers v of | X ϕv (x)g(x)dρ(x)| where the value
is positive (resp. negative).
This implies that:
Z

max hf, giL2 (dρ) = δ max
ϕv (x)g(x)dρ(x), (1)
γ1 (f )6δ v∈V X
with the maximizers f of the first optimization problem above (left-hand side) obtained
as δ times convex combinations of ϕv and −ϕv for maximizers v of the second problem
(right-hand side).
A common difficulty in practice is the hardness of the Frank-Wolfe step, that is, the
optimization problem above over V may be difficult to solve. See Section 3.2, 3.3 and 3.4
for neural networks, where this optimization is usually difficult.
Finitely many observations. When X is finite (or when using the result from Sec-
tion 2.2), the Frank-Wolfe step in Eq. (1) becomes equivalent to, for some vector g ∈ Rn :
n n
1X 1 X
sup gi f (xi ) = δ max
gi ϕv (xi ), (2)
γ1 (f )6δ n i=1
v∈V n
i=1
where the set of solutions of the first problem is in the convex hull of the solutions of the
second problem.
Non-smooth loss functions. In this paper, in our theoretical results, we consider non-
smooth loss functions for which conditional gradient algorithms do not converge in general.
One possibility is to smooth the loss function, as done by Nesterov (2005): an approximation
of ε may be obtained with a smoothness √
error √ constant proportional to 1/ε. By choosing ε
as 1/ t, we obtain a convergence rate of O(1/ t) after t iterations. See also Lan (2013).
Approximate oracles. The conditional gradient algorithm may deal with approximate
oracles; however, what we need in this paper is not the additive errors situations considered
by Jaggi (2013), but multiplicative ones on the computation of the dual norm (similar to
ones derived by Bach (2013) for the regularized problem).
Indeed, in our context, we minimize a function J(f ) on f ∈ L2 (dρ) over a norm ball
{γ1 (f ) 6 δ}. A multiplicative approximate oracle outputs for any g ∈ L2 (dρ), a vector
fˆ ∈ L2 (dρ) such that γ1 (fˆ) = 1, and
hfˆ, gi 6 max hf, gi 6 κ hfˆ, gi,

γ1 (f )61
for a fixed κ > 1. In Appendix B, we propose a modification of the conditional gradient

algorithm that converges to a certain h ∈ L2 (dρ) such that γ1 (h) 6 δ and for which
inf γ1 (f )6δ J(f ) 6 J(h) 6 inf γ1 (f )6δ/κ J(f ).
Such approximate oracles are not available in general, because they require uniform
bounds over all possible values of g ∈ L2 (dρ). In Section 5.5, we show that a weaker form
of oracle is sufficient to preserve our generalization bounds from Section 5.
10
Approximation of any function by a finite number of basis functions. The Frank-

Wolfe algorithm may be applied in the function space F1 with J(f ) = 12 E[(f (x)−g(x))2 ], we
get a function ft , supported by t basis functions such that E[(ft (x) − g(x))2 ] = O(γ(g)2 /t).
Hence, any function in F1 may be approximated with averaged error ε with t = O([γ(g)/ε]2 )
units. Note that the conditional gradient algorithm is one among many ways to obtain such
approximation with ε−2 units (Barron, 1993; Kurkova and Sanguineti, 2001; Mhaskar, 2004).
See Section 4.1 for a (slightly) better dependence on ε for convex neural networks.
3. Neural Networks with Non-decreasing Positively Homogeneous

Activation Functions
In this paper, we focus on a specific family of basis functions, that is, of the form
x 7→ σ(w> x + b),
for specific activation functions σ. We assume that σ is non-decreasing and positively

homogeneous of some integer degree, i.e., it is equal to σ(u) = (u)α+ , for some α ∈ {0, 1, . . .}.
We focus on these functions for several reasons:
– Since they are not polynomials, linear combinations of these functions can approxi-
mate any measurable function (Leshno et al., 1993).
– By homogeneity, they are invariant by a change of scale of the data; indeed, if all
observations x are multiplied by a constant, we may simply change the measure µ
defining the expansion of f by the appropriate constant to obtain exactly the same
function. This allows us to study functions defined on the unit-sphere.
– The special case α = 1, often referred to as the rectified linear unit, has seen consider-
able recent empirical success (Nair and Hinton, 2010; Krizhevsky et al., 2012), while
the case α = 0 (hard thresholds) has some historical importance (Rosenblatt, 1958).
The goal of this section is to specialize the results from Section 2 to this particular case and
show that the “Frank-Wolfe” steps have simple geometric interpretations.
We first show that the positive homogeneity of the activation functions allows to transfer
the problem to a unit sphere.
Boundedness assumptions. For the theoretical analysis, we assume that our data in-
puts x ∈ Rd are almost surely bounded by R in `q -norm, for some q ∈ [2, ∞] (typically
q = 2 and q = ∞). We then build the augmented variable z ∈ Rd+1 as√z = (x> , R)> ∈ Rd+1
by appending the constant R to x ∈ Rd . We therefore have kzkq 6 2R. By defining the
vector v = (w> , b/R)> ∈ Rd+1 , we have:
ϕv (x) = σ(w> x + b) = σ(v > z) = (v > z)α+ ,
which now becomes a function of z ∈ Rd+1 .

Without loss of generality (and by homogeneity of σ), we may assume that the `p -norm
of each vector v is equal to 1/R, that is V will be the (1/R)-sphere for the `p -norm, where
1/p + 1/q = 1 (and thus p ∈ [1, 2], with corresponding typical values p = 2 and p = 1).
11
Bach
This implies by Hölder’s inequality that ϕv (x)2 6 2α . Moreover this leads to functions
in F1 that are bounded everywhere, that is, ∀f ∈ F1 , f (x)2 6 2α γ1 (f )2 . Note that the
functions in F1 are also Lipschitz-continuous for α > 1.
√ Since all `p -norms (for p ∈ [1, 2]) are equivalent to each other with constants of at most
d with respect to the `2 -norm, all the spaces F1 defined above are equal, but the norms
γ1 are of course different and they differ by a constant of at most dα/2 —this can be seen by
computing the dual norms like in Eq. (2) or Eq. (1).
Homogeneous reformulation. In our study of approximation properties, it will be use-
ful to consider the the space of functionR G1 defined for z in the unit sphere Sd ⊂ Rd+1 of
>
the Euclidean norm, such that g(z) = Sd σ(v z)dµ(v), with the norm γ1 (g) defined as the
infimum of |µ|(Sd ) over all decompositions of g. Note the slight overloading of notations for
γ1 (for norms in G1 and F1 ) which should not cause any confusion.
In order to prove the approximation properties (with unspecified constants depending
only on d), we may assume that p = 2, since the norms k · kp for p ∈ [1, ∞] are equivalent
to k · k2 with a constant that grows at most as dα/2 with respect to the `2 -norm. We thus
focus on the `2 -norm in all proofs in Section 4.
We may go from G1 (a space of real-valued functions defined on the unit `2 -sphere in
d + 1 dimensions) to the space F1 (a space of real-valued functions defined on the ball of
radius R for the `2 -norm) as follows (this corresponds to sending a ball in Rd into a spherical
cap in dimension d + 1, as illustrated in Figure 2).
2 α/2 1

x
kxk
– Given g ∈ G1 , we define f ∈ F1 , with f (x) = R22 + 1 g p . If
kxk 2
2 + R 2 R
g may be represented as Sd σ(v > z)dµ(v), then the function f that we have defined
R
may be represented as
kxk2 α/2 Z α
2 > 1 x
f (x) = 2
+1 v p dµ(v)
R Sd kxk22 + R2 R +
Z α
> x/R
Z
= v dµ(v) = σ(w> x + b)dµ(Rw, b),
S d 1 + Sd
that is γ1 (f ) 6 γ1 (g), because we have assumed that (w> , b/R)> is on the (1/R)-
sphere.
– Conversely, given f ∈ F1 , for z = (t> , a)> ∈ Sd , we define g(z) = g(t, a) = f ( Rt α

a )a ,
> >
which we define as such on the set of z = (t , a) ∈ R p d × R (of unit norm)
1 2
√ such that
2
a > 2 . Since we always assume kxk2 6 R, we have kxk2 + R 6 2R, and the
√
value of g(z, a) for a > √1is enough to recover f from the formula above.
2
√
On that portion {a > 1/ 2} of the sphere Sd , this function exactly inherits the
differentiability properties of f . That is, (a) if f is bounded by 1 and f is (1/R)-
Lipschitz-continuous, then g is Lipschitz-continuous with a constant that only depends
on d and α and (b), if all derivatives of order less than k are bounded by R−k , then
all derivatives of the same order of g are bounded by a constant that only depends
on d and α. Precise notions of differentiability may be defined on the sphere, using
12
R z
x
Figure 2: Sending a ball to a spherical cap.
the manifold structure (see, e.g., Absil et al., 2009) or through polar coordinates (see,
e.g., Atkinson and Han, 2012, Chapter 3). See these references for more details.
The only remaining important aspect is to define g on the entire sphere, so that (a) its
regularity constants are controlled by a constant times the ones on the portion of the
sphere where it is already defined, (b) g is either even or odd (this will be important
in Section 4). Ensuring that the regularity conditions can be met is classical when
extending to the full sphere (see, e.g., Whitney, 1934). Ensuring that the function
may be chosen as odd or even may be obtained by multiplying√the function g by an
infinitely differentiable function which is equal to one for a > 1/ 2 and zero for a 6 0,
and extending by −g or g on the hemi-sphere a < 0.
In summary, we may consider in Section 4 functions defined on the sphere, which are
much easier to analyze. In the rest of the section, we specialize some of the general con-
cepts reviewed in Section 2 to our neural network setting with specific activation functions,
namely, in terms of corresponding kernel functions and geometric reformulations of the
Frank-Wolfe steps.
3.1 Corresponding Positive-definite Kernels

In this section, we consider the `2 -norm on the input weight vectors w (that is p = 2). We
may compute for x, x0 ∈ Rd the kernels defined in Section 2.3:
kα (x, x0 ) = E[(w> x + b)α+ (w> x0 + b)α+ ],
Sd , and 0 d+1
for (Rw, b) distributed uniformly on the unit `2 -sphere
r r x, x ∈ R . Given the
x> x0 kxk22 kx0 k22
angle ϕ ∈ [0, π] defined through + 1 = (cos ϕ) + 1 + 1, we have explicit
R2 R2 R2
expressions (Le Roux and Bengio, 2007; Cho and Saul, 2009):
1
k0 (z, z 0 ) = (π − ϕ)
2π
q q
kxk22 kx0 k22
0 R2
+1 R2
+1
k1 (z, z ) = ((π − ϕ) cos ϕ + sin ϕ)
2(d + 1)π
13
Bach
kxk22 kx0 k22

R2
+1 R2
+1
k2 (z, z 0 ) = (3 sin ϕ cos ϕ + (π − ϕ)(1 + 2 cos2 ϕ)).
2π[(d + 1)2 + 2(d + 1)]
There are key differences and similarities between the RKHS F2 and our space of functions
F1 . The RKHS is smaller than F1 (i.e., the norm in the RKHS is larger than the norm in
F1 ); this implies that approximation properties of the RKHS are transferred to F1 . In fact,
our proofs rely on this fact.
However, the RKHS norm does not lead to any adaptivity, while the function space F1
does (see more details in Section 5). This may come as a paradox: both the RKHS F2 and
F1 have similar properties, but one is adaptive while the other oneR is not. A key intuitive
differenceR is as follows: given a function f expressed as f (x) = RV ϕv (x)p(v)dτ (v), then
γ1 (f ) = V |p(v)|dτ (v), while the squared RKHS norm is γ2 (f )2 = V |p(v)|2 dτ (v). For the
L1 -norm, the measure p(v)dτ (v) may tend to a singular distribution with a bounded norm,
while this is not true for the L2 -norm. For example, the function (w> x + b)α+ is in F1 , while
it is not in F2 in general.
3.2 Incremental Optimization Problem for α = 0

We consider the problem in Eq. (2) for the special case α = 0. For z1 , . . . , zn ∈ Rd+1 and
a vector y ∈ Rn , the goal is to solve (as well as the corresponding problem with y replaced
by −y):
Xn X X
max yi 1v> zi >0 = max |yi |1v> zi >0 − |yi |1v> zi >0 ,
v∈Rd+1 v∈Rd+1
i=1 i∈I+ i∈I−
where I+ = {i, yi > 0} and I− = {i, yi < 0}. As outlined by Bengio et al. (2006), this
is equivalent to finding an hyperplane parameterized by v that minimizes a weighted mis-
classification rate (when doing linear classification). Note that the norm of v has no effect.
NP-hardness. This problemP is NP-hard in general. Indeed, if we assume that all yi
are equal to −1 or 1 and with ni=1 yi = 0, then we have Pna balanced binary classification
n
problem (we need to assume n even). The quantity i=1 yi 1v > zi >0 is then 2 (1 − 2e)
where e is the corresponding classification error for a problem of classifying at positive
(resp. negative) the examples in I+ (resp. I− ) by thresholding the linear classifier v > z.
Guruswami and Raghavendra (2009) showed that for all (ε, δ), it is NP-hard to distinguish
between instances (i.e., configurations of points xi ), where a halfspace with classification
error at most ε exists, and instances where all half-spaces have an error of at least 1/2 − δ.
d+1 such that
Thus, it is NP-hard to distinguish between instances where there exists
P n n d+1
Pnv ∈ R
i=1 yi 1v > zi >0 > 2 (1 − 2ε) and instances where for all v ∈ R Pn , i=1 yi 1v> zin>0 6 nδ.
Thus, it is NP-hard to distinguish instances where maxv∈Rd+1 i=1 yi 1v> zi >0 > 2 (1 − 2ε)
and ones where it is less than n2 δ. Since this is valid for all δ and ε, this rules out a
constant-factor approximation.
Convex relaxation. Given linear binary classification problems, there are several algo-
rithms to approximately find a good half-space. These are based on using convex surrogates
(such as the hinge loss or the logistic loss). Although some theoretical results do exist regard-
ing the classification performance of estimators obtained from convex surrogates (Bartlett
et al., 2006), they do not apply in the context of linear classification.
14
3.3 Incremental Optimization Problem for α = 1

We consider the problem in Eq. (2) for the special case α = 1. For z1 , . . . , zn ∈ Rd+1 and
a vector y ∈ Rn , the goal is to solve (as well as the corresponding problem with y replaced
by −y):
Xn X X
max yi (v > zi )+ = max (v > |yi |zi )+ − (v > |yi |zi )+ ,
kvkp 61 kvkp 61
i=1 i∈I+ i∈I−
where I+ = {i, yi > 0} and I− = {i, yi < 0}. We have, with ti = |yi |zi ∈ Rd+1 , using convex
duality:
n
X X X
max yi (v > zi )+ = max (v > ti )+ − (v > ti )+
kvkp 61 kvkp 61
i=1 i∈I+ i∈I−
X X
= max max bi v > ti − max bi v > ti
kvkp 61 bi ∈[0,1] bi ∈[0,1]
i∈I+ i∈I−
= max max min v [T+> b+ − T−> b− ]

>
b+ ∈[0,1] I+ kvkp 61 b− ∈[0,1]I−
= max min max v > [T+> b+ − T−> b− ] by Fenchel duality,

b+ ∈[0,1]I+ b− ∈[0,1]I− kvkp 61
= max min kT+> b+ − T−> b− kq ,

I+ I−
b+ ∈[0,1] b− ∈[0,1]
where T+ ∈ Rn+ ×d has rows ti , i ∈ I+ and T− ∈ Rn− ×d has rows ti , i ∈ I− , with v ∈

arg maxkvkp 61 v > (T+> b+ − T−> b− ). The problem thus becomes
max min kT+> b+ − T−> b− kq .

b+ ∈[0,1]n+ b− ∈[0,1]n−
Pn >z
For the problem of maximizing | i=1 yi (v i )+ |, then this corresponds to

max max min kT+> b+ − T−> b− kq , max min kT+> b+ − T−> b− kq .
b+ ∈[0,1]n+ b− ∈[0,1]n− b− ∈[0,1]n− b+ ∈[0,1]n+
This is exactly the Hausdorff distance between the two convex sets {T+> b+ , b+ ∈ [0, 1]n+ }
and {T−> b− , b− ∈ [0, 1]n− } (referred to as zonotopes, see below).
Given the pair (b+ , b− ) achieving the Hausdorff distance, then we may compute the
optimal v as v = arg maxkvkp 61 v > (T+> b+ − T−> b− ). Note this has not changed the problem
at all, since it is equivalent. It is still NP-hard in general (König, 2014). But we now have a
geometric interpretation with potential approximation algorithms. See below and Section 6.
Zonotopes. A zonotope A is the Minkowski sum of a finite number of segments from the
origin, that is, of the form
r
nX o
A = [0, t1 ] + · · · + [0, tr ] = bi ti , b ∈ [0, 1]r ,
i=1
15
Bach
t1 + t2 + t3
t1
t3
t2
0 0 0 0
Figure 3: Two zonotopes in two dimensions: (left) vectors, and (right) their Minkowski sum
(represented as a polygone).
0
0 0
Figure 4: Left: two zonotopes (with their generating segments) and the segments achieving
the two sides of the Haussdorf distance. Right: approximation by ellipsoids.
for some vectors ti , i = 1, . . . , r (Bolker, 1969). See an illustration in Figure 3. They

appear in several areas of computer science (Edelsbrunner, 1987; Guibas et al., 2003) and
mathematics (Bolker, 1969; Bourgain et al., 1989). In machine learning, they appear natu-
rally as the affine projection of a hypercube; in particular, when using a higher-dimensional
distributed representation of points in Rd with elements in [0, 1]r , where r is larger than
d (see, e.g., Hinton and Ghahramani, 1997), the underlying polytope that is modelled in Rd
happens to be a zonotope.
In our context, the two convex sets {T+> b+ , b+ ∈ [0, 1]n+ } and {T−> b− , b− ∈ [0, 1]n− }
defined above are thus zonotopes. See an illustration of the Hausdorff distance computation
in Figure 4 (middle plot), which is the core computational problem for α = 1.
Approximation by ellipsoids. Centrally symmetric convex polytopes (w.l.o.g. centered

around zero) may be approximated by ellipsoids. In our set-up, we could use the minimum
volume enclosing ellipsoid (see, e.g. Barvinok, 2002), which can be computed exactly when
the polytope is given through its vertices, or up to a constant factor when the polytope is
such that quadratic functions may be optimized with a constant factor approximation. For
zonotopes, the standard semi-definite relaxation of Nesterov (1998) leads to such constant-
factor approximations, and thus the minimum volume inscribed ellipsoid√ may be computed
up to a constant. Given standard results (see, e.g. Barvinok, 2002), a (1/ d)-scaled version
of the ellipsoid is inscribed in this polytope, and thus √the ellipsoid is a provably good
approximation of the zonotope with a factor scaling as d. However, the approximation
16
ratio is not good enough to get any relevant bound for our purpose (see Section 5.5), as for
computing the Haussdorff distance, we care about potentially vanishing differences that are
swamped by constant factor approximations.
Nevertheless, the ellipsoid approximation may prove useful in practice, in particular
because the `2 -Haussdorff distance between two ellipsoids may be computed in polynomial
time (see Appendix E).
NP-hardness. Given the reduction of the case α = 1 (rectified linear units) to α = 0
(exact thresholds) (Livni et al., 2014), the incremental problem is also NP-hard, so as ob-
taining a constant-factor approximation. However, this does not rule out convex relaxations
with non-constant approximation ratios (see Section 6 for more details).
3.4 Incremental Optimization Problem for α > 2

We consider the problem in Eq. (2) for the remaining cases α > 2. For z1 , . . . , zn ∈ Rd+1
and a vector y ∈ Rn , the goal is to solve (as well as the corresponding problem with y
replaced by −y):
n
1X X 1 X 1
max yi (v > zi )α+ = max (v > |yi |1/α zi )α+ − (v > |yi |1/α zi )α+ ,
kvkp 61 α kvkp 61 α α
i=1 i∈I+ i∈I−
where I+ = {i, yi > 0} and I− = {i, yi < 0}. We have, with ti = |yi |1/α zi ∈ Rd+1 , and
β ∈ (1, 2] defined by 1/β + 1/α = 1 (we use the fact that the function u 7→ uα /α and
v 7→ v β /β are Fenchel-dual to each other):
n
1X X 1 X 1
max yi (v > zi )α+ = max (v > ti )α+ − (v > ti )α+
kvkp 61 α kvkp 61 α α
i=1 i∈I+ i∈I−
n 1 o X n 1 o
max bi vi> ti − bβi − max bi v > ti − bβi
X
= max
kvkp 61 bi >0 β bi >0 β
i∈I+ i∈I−
1 1
= max min max v > [T+> b+ − T−> b− ] − kb+ kββ + kb− kββ
I
b+ ∈R++
I
b− ∈R+− kvkp 61 β β
by Fenchel duality,
1 1
= max min kT+> b+ − T−> b− kq − kb+ kββ + kb− kββ , (3)
I+
b+ ∈[0,1] b− ∈[0,1] I− β β
where T+ ∈ Rn+ ×d has rows ti , i ∈ I+ and T− ∈ Rn− ×d has rows ti , i ∈ I− , with v ∈

arg maxkvkp 61 (T+> b+ − T−> b− )> v. Contrary to the case α = 1, we do not obtain exactly
a formulation as a Hausdorff distance. However, if we consider the convex sets Kλ+ =
{T+> b+ , b+ > 0, kb+ kβ 6 λ} and Kµ− = {T−> b− , b− > 0, kb− kβ 6 µ}, then, a solution of
Eq. (3) may be obtained from Hausdorff distance computations between Kλ+ and Kµ− , for
certain λ and µ.
Note that, while for α = 1 we can use the identity 2u+ = u + |u| to replace the
rectified linear unit by the absolute value and obtain the same function space, this is not
possible for α = 2, as (u+ )2 and u2 do not differ by a linear function. This implies that
17
Bach
the results from Livni et al. (2014), which state that for the quadratic activation function,
the incremental problems is equivalent to an eigendecomposition (and hence solvable in
polynomial time), do not apply.
4. Approximation Properties
In this section, we consider the approximation properties of the set F1 of functions defined
on Rd . As mentioned earlier, the norm used to penalize input weights w or v is irrelevant
for approximation properties as all norms are equivalent. Therefore, we focus on the case
q = p = 2 and `2 -norm constraints.
Because we consider homogeneous activation functions, we start by studying the set
G1 of functions defined on the unit `2 -sphere Sd ⊂ Rd+1 . We denote by τd the uniform
probability measureR on > Sd . The set G1 is defined as the set of functions on the sphere
such thatR g(z) = Sd σ(v z)p(z)dτd (z), with the norm γ1 (g) equal to the smallest possible
value of Sd |p(z)|dτd (z). We
R may also define the corresponding squared RKHS norm by the
smallest possible value of Sd |p(z)|2 dτd (z), with the corresponding RKHS G2 .
In this section, we first consider approximation properties of functions in G1 by a finite
number of neurons (only for α = 1). We then study approximation properties of functions on
the sphere by functions in G1 . It turns out that all our results are based on the approximation
properties of the corresponding RKHS G2 : we give sufficient conditions for being in G2 , and
then approximation bounds for functions which are not in G2 . Finally we transfer these to
the spaces F1 and F2 , and consider in particular functions which only depend on projections
on a low-dimensional subspace, for which the properties of G1 and G2 (and of F1 and F2 )
differ. This property is key to obtaining generalization bounds that show adaptivity to
linear structures in the prediction functions (as done in Section 5).
Approximation properties of neural networks with finitely many neurons have been
studied extensively (see, e.g., Petrushev, 1998; Pinkus, 1999; Makovoz, 1998; Burger and
Neubauer, 2001). In Section 4.7, we relate our new results to existing work from the
literature on approximation theory, by showing that our results provide an explicit control
of the various weight vectors which are needed for bounding the estimation error in Section 5.
4.1 Approximation by a Finite Number of Basis Functions

A key quantity that drives the approximability by a finite number of neurons is the variation
norm γ1 (g). As shown in Section 2.5, any function g such that γ1 (g) is finite, may be
approximated in L2 (Sd )-norm with error ε with n = O(γ1 (g)2 ε−2 ) units. For α = 1 (rectified
linear units), we may improve the dependence in ε, through the link with zonoids and
zonotopes, as we now present.
If we decompose the signed measureR µ as µ = µ+ −µ− where µ+ and µ− are positive mea-
sures, then, for g ∈ G1 , we have g(z) = Sd (v > z)+ dµ+ (v)− Sd (v > z)+ dµ− (v) = g+ (z)−g− (z),
R
which is a decomposition of g as a difference of positively homogenous convex functions.

Positively homogenous convex functions h may be written as the support function of
a compact convex set K (Rockafellar, 1997), that is, h(z) = maxy∈K y > z, and the set K
characterizes the function h. The functions g+ and g− defined above are not any convex
positively homogeneous functions, as we now describe.
18
µ+ (v) = ri=1 ηi δ(v −vi )

P
If the measure µ+ is supported by finitely many points, that is,
with η > 0, then g+ (z) = i=1 ηi (vi> z)+ = ti=1 (ηi vi> z)+ = ti=1P
Pt
(t>
P P
i z)+ for ti = ηi vi .
Thus the corresponding set K+ is the zonotope [0, t1 ] + · · · + [0, tr ] = { ri=1 bi ti , b ∈ [0, 1]r }
already defined in Section 3.3. Thus the functions g+ ∈ G1 and g− ∈ G1 for finitely supported
measures µ are support functions of zonotopes.
When the measure µ is not constrained to have finite support, then the sets K+ and K−
are limits of zonotopes, and thus, by definition, zonoids (Bolker, 1969), and thus functions
in G1 are differences of support functions of zonoids. Zonoids are a well-studied set of con-
vex bodies. They are centrally symmetric, and in two dimensions, all centrally symmetric
compact convexs sets are (up to translation) zonoids, which is not true in higher dimen-
sions (Bolker, 1969). Moreover, the problem of approximating a zonoid by a zonotope with
a small number of segments (Bourgain et al., 1989; Matoušek, 1996) is essentially equiva-
lent to the approximation of a function g by finitely many neurons. The number of neurons
directly depends on the norm γ1 , as we now show.
Proposition 1 (Number of units - α = 1) Let ε ∈ (0, 1/2). For any function g in G1 ,

there exists a measure µ supported on at most r points in V, so that for all z ∈ Sd . |g(z) −
> −2d/(d+3) , for some constant C(d) that depends
R
Sd (v z)+ dµ(v)| 6 εγ1 (g), with r 6 C(d)ε
only on d.
Proof Without loss of generality, we assume γ(g) = 1. It is shown by Matoušek (1996)

that for any probability measure µ (positive and with finite mass) on the sphere Sd , there
exists a set of r points v1 , . . . , vr , so that for all z ∈ Sd ,
Z r
> 1 X >
|v z|dµ(v) − |vi z| 6 ε, (4)

Sd r
i=1
with r 6 C(d)ε−2+6/(d+3) = C(d)ε−2d/(d+3) , for some constant C(d) that depends only on
d. We may then simply write
µ+ (Sd ) dµ+ (v) µ− (Sd ) dµ− (v)

Z Z Z Z
1
g(z) = (v > z)+ dµ(v) = (v > z)dµ(v)+ |v > z| − |v > z| ,
Sd 2 Sd 2 Sd µ+ (Sd ) 2 Sd µ− (Sd )
and approximate the last two terms with error εµ± (Sd ) with r terms, leading to an approxi-
mation of εµ+ (Sd ) + εµ− (Sd ) = εγ1 (g) = ε, with a remainder that is a linear function q > z of
z, with kqk2 6 1. We may then simply add two extra units with vectors q/kqk2 and weights
−kqk2 and kqk2 . We thus obtain, with 2r + 2 units, the desired approximation result.
Note that Bourgain et al. (1989, Theorem 6.5) showed that the scaling in ε in Eq. (4)
is not improvable, if the measure is allowed to have non equal weights on all points and the
proof relies on the non-approximability of the Euclidean ball by centered zonotopes. This
results does not apply here, because we may have different weights µ− (Sd ) and µ+ (Sd ).
Note that the proposition above is slightly improved in terms of the scaling of the num-
ber of neurons with respect to the approximation error ε (improved exponent), compared
to conditional gradient bounds (Barron, 1993; Kurkova and Sanguineti, 2001). Indeed, the
19
Bach
simple use of conditional gradient leads to r 6 ε−2 γ1 (g)2 , with a better constant (indepen-
dent of d) but a worse scaling in ε—also with a result in L2 (Sd )-norm and not uniformly on
the ball {kxkq 6 R}. Note also that the conditional gradient algorithm gives a construc-
tive way of building the measure. Moreover, the proposition above is related to the result
from Makovoz (1998, Theorem 2), which applies for α = 0 but with a number of neurons
growing as ε−2d/(d+1) , or to the one of Burger and Neubauer (2001, Example 3.1), which
applies to a piecewise affine sigmoidal function but with a number of neurons growing as
ε−2(d+1)/(d+3) (both slightly worse than ours).
Finally, the number of neurons needed to express a function with a bound on the γ2 -
norm can be estimated from general results on approximating reproducing kernel Hilbert
space described in Section 2.3, whose kernel can be expressed as an expectation. Indeed,
Bach (2017) shows that with k neurons, one can approximate a function in F2 with unit
γ2 -norm with an error measured in L2 of ε = k −(d+3)/(2d) . When inverting the relationship
between k and ε, we get a number of neurons scaling as ε−2d/(d+3) , which is the same as in
Prop. 1 but with an error in L2 -norm instead of L∞ -norm.
4.2 Sufficient Conditions for Finite Variation

In this section and the next one, we study more precisely the RKHS G2 (and thus obtain
similar results for G1 ⊃ G2 ). The kernel k(x, y) = Sd (v x)+ (v > y)+ dτd (v) defined on the
>
R
sphere Sd belongs to the family of dot-product kernels (Smola et al., 2001) that only depends
on the dot-product x> y, although in our situation, the function is not particularly simple
(see formulas in Section 3.1). The analysis of these kernels is similar to one of translation-
invariant kernels; for d = 1, i.e., on the 2-dimensional sphere, it is done through Fourier
series; while for d > 1, spherical harmonics have to be used as the expansion of functions
in series of spherical harmonics make the computation of the RKHS norm explicit (see a
review of spherical harmonics in Appendix D.1 with several references therein). Since the
calculus is tedious, all proofs are put in appendices, and we only present here the main
results. In this section, we provide simple sufficient conditions for belonging to G2 (and
hence G1 ) based on the existence and boundedness of derivatives, while in the next section,
we show how any Lipschitz-function may be approximated by functions in G2 (and hence
G1 ) with precise control of the norm of the approximating functions.
The derivatives of functions defined on Sd may be defined in several ways, using the
manifold structure (see, e.g., Absil et al., 2009) or through polar coordinates (see, e.g.,
Atkinson and Han, 2012, Chapter 3). For d = 1, the one-dimensional sphere S1 ⊂ R2 may
be parameterized by a single angle and thus the notion of derivatives and the proof of the
following result is simpler and based on Fourier series (see Appendix C.2). For the general
proof based on spherical harmonics, see Appendix D.2.
Proposition 2 (Finite variation on the sphere) Assume that g : Sd → R is such that

all i-th order derivatives exist and are upper-bounded in absolute value by η for i ∈ {0, . . . , s},
where s is an integer such that s > (d − 1)/2 + α + 1. Assume g is even if α is odd (and
vice-versa); then g ∈ G2 and γ2 (g) 6 C(d, α)η, for a constant C(d, α) that depends only on
d and α.
We can make the following observations:
20
– Tightness of conditions: as shown in Appendix D.5, there are functions g, which have
bounded first s derivatives and do not belong to G2 while s 6 d2 + α (at least when
s − α is even). Therefore, when s − α is even, the scaling in (d − 1)/2 + α is optimal.
– Dependence on α: for any d, the higher the α, the stricter the sufficient condition.
Given that the estimation error grows slowly with α (see Section 5.1), low values of
α would be preferred in practice.
– Dependence on d: a key feature of the sufficient condition is the dependence on d, that

is, as d increases the number of derivatives has to increase in d/2—like for Sobolev
spaces in dimension d (Adams and Fournier, 2003). This is another instantiation
of the curse of dimensionality: only very smooth functions in high dimensions are
allowed.
– Special case d = 1, α = 0: differentiable functions on the sphere in R2 , with bounded

derivatives, belong to G2 , and thus all Lipschitz-continuous functions, because Lipschitz-
continuous functions are almost everywhere differentiable with bounded derivative
(Adams and Fournier, 2003).
4.3 Approximation of Lipschitz-continuous Functions

In order to derive generalization bounds for target functions which are not sufficiently
differentiable (and may not be in G2 or G1 ), we need to approximate any Lipschitz-continuous
function, with a function g ∈ G2 with a norm γ2 (g) that will grow as the approximation gets
tighter. We give precise rates in the proposition below. Note the requirement for parity of
the function g. The result below notably shows the density of G1 in uniform norm in the
space of Lipschitz-continuous functions of the given parity, which is already known since
our activation functions are not polynomials (Leshno et al., 1993).
Proposition 3 (Approximation of Lipschitz-continuous functions on the sphere)

For δ greater than a constant depending only on d and α, for any function g : Sd → R such
that for all x, y ∈ Sd , g(x) 6 η and |g(x) − g(y)| 6 ηkx − yk2 , and g is even if α is odd (and
vice-versa), there exists h ∈ G2 , such that γ2 (h) 6 δ and
δ −1/(α+(d−1)/2) δ
sup |h(x) − g(x)| 6 C(d, α)η log .
x∈Sd η η
This proposition is shown in Appendix C.3 for d = 1 (using Fourier series) and in
Appendix D.4 for all d > 1 (using spherical harmonics). We can make the following obser-
vations:
– Dependence in δ and η: as expected, the main term in the error bound (δ/η)−1/(α+(d−1)/2)
is a decreasing function of δ/η, that is when the norm γ2 (h) is allowed to grow, the
approximation gets tighter, and when the Lipschitz constant of g increases, the ap-
proximation is less tight.
– Dependence on d and α: the rate of approximation is increasing in d and α. In

particular the approximation properties are better for low α.
21
Bach
– Special case d = 1 and α = 0: up to the logarithmic term we recover the result of

Prop. 2, that is, the function g is in G2 .
– Tightness: in Appendix D.5, we provide a function which is not in the RKHS and
for which the tightest possible approximation scales as δ −2/(d/2+α−2) . Thus the linear
scaling of the rate as d/2 + α is not improvable (but constants are).
4.4 Linear Functions

In this section, we consider a linear function on Sd , that is g(x) = v > x for a certain v ∈ Sd ,
and compute its norm (or upper-bound thereof) both for G1 and G2 , which is independent
of v and finite. In the following propositions, the notation ≈ means asymptotic equivalents
when d → ∞.
Proposition 4 (Norms of linear functions on the sphere) Assume that g : Sd → R

2dπ
is such g(x) = v > x for a certain v ∈ Sd . If α = 0, then γ1 (g) 6 γ2 (g) = d−1 ≈ 2π. If
d 4π Γ(α/2+d/2+1)
α = 1, then γ1 (g) 6 2, and for all α > 1, γ1 (g) 6 γ2 (g) = d−1 α Γ(α/2)Γ(d/2+1) ≈ Cdα/2 .
We see that for α = 1, the√ γ1 -norm is less than a constant, and is much smaller than
the γ2 -norm (which scales as d). For α > 2, we were not able to derive better bounds for
γ1 (other than the value of γ2 ).
4.5 Functions of Projections

If g(x) = ϕ(w> x) for some unit-norm w ∈ Rd+1 and ϕ a function defined on the real-line,
then the value of the norms γ2 and γ1 differ significantly. Indeed, for γ1 , we may consider
a new variable x̃ ∈ S1 ⊂ R2 , with its first component x̃1 = w> x, and the function g̃(x) =
ϕ(x̃1 ). We may then apply Prop. 2 to g̃ with d = 1. That is, if ϕ is (α+1)-times differentiable
with bounded derivatives, there exists a decomposition g̃(x̃) = S1 µ̃(ṽ)σ(ṽ > x̃)dµ̃, with
R
γ1 (g̃) = |µ̃|(S1 ), which is not increasing in d. If we consider any vector t ∈ Rd+1 which is
orthogonal to w in Rd+1 , then, we may define a measure µ supported in the circle defined by
the two vectors w and t and which is equal to µ̃ on that circle. The total variation of µ is the
one of µ̃ while g can be decomposed using µ and thus γ1 (g) 6 γ1 (g̃). Similarly, Prop. 3 could
also be applied (and will for obtaining generalization bounds), also our reasoning works for
any low-dimensional projections: the dependence on a lower-dimensional projection allows
to reduce smoothness requirements.
However, for the RKHS norm γ2 , this reasoning does not apply. For example, a certain
function ϕ exists, which is s-times differentiable, as shown in Appendix D.5, for s 6 d2 + α
(when s − α is even), and is not in G2 . Thus, given Prop. 2, the dependence on a uni-
dimensional projection does not make a difference regarding the level of smoothness which
is required to belong to G2 .
4.6 From the Unit-sphere Sd to Rd+1

We now extend the results above to functions defined on Rd , to be approximated by functions
in F1 and F2 . More precisely, we first extend Prop. 2 and Prop. 3, and then consider norms
of linear functions and functions of projections.
22
Proposition 5 (Finite variation) Assume that f : Rd → R is such that all i-th order
derivatives exist and are upper-bounded on the ball {kxkq 6 R} by η/Ri for i ∈ {0, . . . , k},
where s is the smallest integer such that s > (d − 1)/2 + α + 1; then f ∈ F2 and γ2 (f ) 6
C(d, α)η, for a constant C(d, α) that depends only on d and α.
Proof By assumption, the function x 7→ f (Rx) has all its derivatives bounded by a con-
stant times η. Moreover, we have defined g(t, a) = f ( Rt α
a )a so that all derivatives are
bounded by η. The result then follows immediately from Prop. 2.
Proposition 6 (Approximation of Lipschitz-continuous functions) For δ larger than

a constant that depends only on d and α, for any function f : Rd → R such that for all x, y
such that kxkq 6 R and kykq 6 R, |f (x)| 6 η and |f (x) − f (y)| 6 ηR−1 kx − ykq , there
exists g ∈ F2 such that γ2 (g) 6 δ and
δ −1/(α+(d−1)/2) δ
sup |f (x) − g(x)| 6 C(d, α)η log .
kxkq 6R η η
Proof With the same reasoning as above, we obtain that g is Lipschitz-continuous with
constant η, we thus get the desired approximation error from Prop. 3.
Linear functions. If f (x) = w> x + b, with kwk2 6 η and b 6 ηR, then for α = 1, it is
straightforward that γ1 (f ) 6 2Rη. Moreover, we have γ2 (f ) ∼ CRη. For other values of
α, we also have γ1 -norms less than a constant (depending only of α) times Rη. The RKHS
norms are bit harder to compute since linear functions for f leads to linear functions for g
only for α = 1.
Functions of projections. If f (x) = ϕ(w> x) where kwk2 6 η and ϕ : R → R is a

function, then the norm of f is the same as the norm of the function ϕ on the interval
[−Rη, Rη], and it thus does not depend on d. This is a consequence of the fact that the
total mass of a Radon measure remains bounded even when the support has measure zero
(which might not be the case for the RKHS defined in Section 2.3). For the RKHS, there
is no such results and it is in general not adaptive.
More generally, if f (x) = Φ(W > x) for W ∈ Rd×s with the largest singular value of W
less than η, and Φ a function from Rs to R, then for kxk2 6 R, we have kW > xk2 6 Rη,
and thus we may apply our results for d = s.
`1 -penalty on input weights (p=1). When using an `1 -penalty on input weights instead
of an `2 -penalty, the results in Prop. 5 and 6 are unchanged (only the constants that
depend on d are changed). Moreover, when kxk∞ 6 1 almost surely, functions of the form
f (x) = ϕ(w> x) where kwk1 6 η and ϕ : R → R is a function, will also inherit from
properties of ϕ (without any dependence on dimension). Similarly, for functions of the
form f (x) = Φ(W > x) for W ∈ Rd×s with all columns of `1 -norm less than η, we have
kW > xk∞ 6 Rη and we can apply the s-dimensional result.
23
Bach
4.7 Related Work

In this section, we show how our results from the previous sections relate to existing work
on neural network approximation theory.
Approximation of Lipschitz-continuous functions with finitely many neurons.
In this section, we only consider the case α = 1, for which we have two approximation
bounds: Prop. 6 which approximates any η-Lipschitz-continuous function by a function
with finite γ1 -norm less than δ and uniform error less than η(δ/η)−2/(d+1) log (δ/η), and
Prop. 1 which shows that a function with γ1 -norm less than δ, may be approximated with
r neurons with uniform error δr−(d+3)/(2d) .
Thus, given r neurons, we get an approximation of the original function with uniform
error
η(δ/η)−2/(d+1) log (δ/η) + δr−(d+3)/(2d) .
We can optimize over δ, and use δ = ηn(d+1)/(2d) , to obtain a uniform approximation bound
proportional to η(log n)n−1/d , for approximating an η-Lipschitz-continuous function with n
neurons.
Approximation by ridge functions. The approximation properties of single hidden
layer neural networks have been studied extensively, where they are often referred to as
“ridge function” approximations. As shown by Pinkus (1999, Corollary 6.10)—based on
a result from Petrushev (1998), the approximation order of n−1/d for the rectified linear
unit was already known, but only in L2 -norm (and without the factor log n), and without
any constraints on the input and output weights. In this paper, we provide an explicit
control of the various weights, which is needed for computing estimation errors. Moreover,
while the two proof techniques use spherical harmonics, the proof of Petrushev (1998)
relies on quadrature formulas for the associated Legendre polynomials, while ours relies on
the relationship with the associated positive definite kernels, is significantly simpler, and
offers additional insights into the problem (relationship with convex neural networks and
zonoids). Maiorov (2006, Theorem 2.3) also derives a similar result, but in L2 -norm (rather
than uniform norm), and for sigmoidal activation functions (which are bounded). Note
finally, that the order O(n−1/d ) cannot be improved (DeVore et al., 1989, Theorem 4.2).
Also, Maiorov and Meir (2000, Theorem 5) derive similar upper and lower bounds based on
a random sampling argument which is close to using random features in the RKHS setting
described in Section 2.3.
Relationship to hardness results for Boolean-valued functions. In this paper, we
consider a particular view of the curse of dimensionality and ways of circumventing it, that
is, our distribution over inputs is arbitrary, but our aim is to approximate a real-valued
function. Thus, all hardness results depending on functions with values in {0, 1} do not
apply there directly—see, e.g., Shalev-Shwartz and Ben-David (2014, Chapter 20), for the
need of exponentially many hidden units for approximating most of the functions from
{0, 1}d to {0, 1}.
Our approximation bounds show that, without any assumption beyond Lipschitz-continuity
of the target function, it sufficient to have a number of hidden units which is still exponen-
tial in dimension (hence we also suffer from the curse of dimensionality), but a soon as the
target function depends on linear low-dimensional structure, then we lose this exponential
24
dependence. It would be interesting to study an extension to {0, 1}-valued functions, and

also to relate our results to the number of linear regions delimited by neural networks with
rectified linear units (Montufar et al., 2014).
5. Generalization Bounds
Our goal is to derive the generalization bounds outlined in Section 2.4 for neural networks
with a single hidden layer. The main results that we obtain are summarized in Table 2 and
show adaptivity to assumptions that avoid the curse of dimensionality.
More precisely, given some distribution over the pairs (x, y) ∈ X × Y, a loss function
` : Y × R → R, our aim is to find a function f : Rd → R such that J(f ) = E[`(y, f (x))]
is small, given some i.i.d. observations (xi , yi ), i = 1, . . . , n. We consider the empirical risk
minimization framework over a space of functions F, equipped with a norm γ (in our situ-
ations, F1 and F2 , equipped with γ1 or γ2 ). The empirical risk J(f ˆ ) = 1 Pn `(yi , f (xi )),
n i=1
is minimized by constraining f to be in the ball F δ = {f ∈ F, γ(f ) 6 δ}.
We assume that almost surely,√ kxkq 6 R, that for all y the function u 7→ `(y, u) is
G-Lipschitz-continuous on {|u| 6 2δ}, √ and that almost surely, `(y, 0) 6 Gδ. As before z
denotes z = (x> , R)> so that kzkq 6 2R. This corresponds to the following examples:
– Logistic regression and support vector machines: we have G = 1.
√ 2
– Least-squares regression: we take G = max { 2δ + kyk∞ , kyk
√ ∞ }.
2δ
Approximation errors inf f ∈F δ J(f ) − inf f ∈F J(f ) will be obtained from the approxima-
tion results from Section 4 by assuming that the optimal target function f∗ has a specific
form. Indeed, we have:
n o
inf J(f ) − J(f∗ ) 6 G inf sup |f (x) − f∗ (x)| .
f ∈F δ f ∈F δ kxkq 6R
ˆ ) − J(f )| using Rademacher complex-

We now deal with estimation errors supf ∈F δ |J(f
ities.
5.1 Estimation Errors and Rademacher Complexity

The following proposition bounds the uniform deviation between J and its empirical coun-
ˆ This result is standard (see, e.g., Koltchinskii, 2001; Bartlett and Mendelson,
terpart J.
2003) and may be extended in bounds that hold with high-probability.
Proposition 7 (Uniform deviations) We have the following bound on the expected uni-
form deviation:
ˆ Gδ
E sup |J(f ) − J(f )| 6 4 √ C(p, d, α),
γ1 (f )6δ n
with the following constants:
√α
p
– for α > 1, C(p, d, α) 6 α 2 log(d + 1) for p = 1 and C(p, d, α) 6 p−1
for p ∈ (1, 2]
√
– for α = 0, C(p, d, α) 6 C d + 1, where C is a universal constant.
25
Bach
Proof We use the standard framework of Rademacher complexities and get:
ˆ )|
E sup |J(f ) − J(f
γ1 (f )6δ
n
1 X
6 2E sup τi `(yi , f (xi )) using Rademacher random variables τi ,
γ1 (f )6δ n i=1
n n
1 X 1 X
6 2E sup τi `(yi , 0) + 2E sup
τi [`(yi , f (xi )) − `(yi , 0)]
γ1 (f )6δ n i=1 γ1 (f )6δ n i=1
X n
Gδ 1
6 2 √ + 2GE sup τi f (xi ) using the Lipschitz-continuity of the loss,
n γ(f )6δ n i=1
1 n
X
Gδ > α

6 2 √ + 2GδE sup τi (v zi )+ using Eq. (2).
n kvkp 61/R n i=1
We then take different routes for α > 1 and α = 0.

For α > 1, we have the upper-bound
1 n
X
Gδ
ˆ )| 6 2 √ τi v > zi

E sup |J(f ) − J(f + 2GδαE sup
γ1 (f )6δ n kvkp 61/R n i=1
using the α-Lipschitz-cont. of (·)α+ on [−1, 1],
Xn
Gδ Gαδ
6 2√ + 2 E τi i .
z
n Rn
q i=1
From Kakade et al. (2009), we get the following bounds on Rademacher complexities:
Pn √ √ √
– If p ∈ (1, 2], then q ∈ [2, ∞), and Ek √1 R
i=1 τi zi kq 6 q − 1R n = p−1
n
Pn √ p
– If p = 1, then q = ∞, and Ek i=1 τi zi kq 6 R n 2 log(d + 1).
Pn √
Overall, we have Ek i=1 τi zi kq 6 nRC(p, d) with C(p, d) defined above, and thus
Gδ
ˆ )| 6 2 √ Gδα
E sup |J(f ) − J(f (1 + αC(p, d)) 6 4 √ C(p, d).
γ(f )6δ n n
For α = 0, we can simply go through the VC-dimension of half-hyperplanes, which is

1 Pn
equal to d, and Theorem 6 from Bartlett and Mendelson (2003), that shows that E supv∈Rd+1 n i=1 τi 1v> zi 6
√
d+1
C √
n
where C is a universal constant.
,
Note that using standard results from Rademacher complexities, we have, r
with proba-
bility greater than 1 − u, sup |J(f ) − J(f ˆ )| + 2Gδ
ˆ )| 6 E sup |J(f ) − J(f √
2
log .
γ1 (f )6δ γ1 (f )6δ n u
26
5.2 Generalization Bounds for `2 -norm Constraints on Input Weights (p = 2)

We now provide generalization bounds for the minimizer of the empirical risk given the
contraint that γ1 (f ) 6 δ for a well chosen δ, that will depend on the assumptions regarding
the target function f∗ , listed in Section 1. In this section, we consider an `2 -norm on input
weights w, while in the next section, we consider the `1 -norm. The two situations are
summarized and compared in Table 2, where we consider that kxk√∞ 6 r almost surely,
which implies that our bound R will depend on dimension as R 6 r d.
Our generalization bounds are expected values of the excess expected risk for a our
estimator (where the expectation is taken over the data).
Affine functions. We assume f∗ (x) = w> x + b, with kwk2 6 η and |b| 6 Rη. Then,
as seen in Section 4.6, f∗ ∈ F1 with γ1 (f∗ ) 6 C(α)ηR (the constant is independent of d
because we approximate an affine function). From Prop. 7, we thus get a generalization
bound proportional to GRη√
n
times a constant (that may depend on α), which is the same
as assuming directly that we optimize over linear predictors only. The chosen δ is then a
constant times Rη, and does not grow with n, like in parametric estimation (although we
do use a non-parametric estimation procedure).
Pk >
Projection pursuit. We assume f∗ (x) = j=1 fj (wj x), with kwj k2 6 η and each
fj bounded by ηR and 1-Lipschitz continuous. From Prop. 6, we may approach each
x 7→ fj (wj> x) by a function with γ1 -norm less than δηR and uniform approximation
C(α)ηRδ −1/α log δ. This leads to a total approximation error of kC(α)GηRδ −1/α log δ for a
norm less than kδηR (the constant is independent of d because we approximate a function
of one-dimensional projection).
For α > 1, from Prop. 7, the estimation error is kGRηδ √
n
, with an overall bound of
C(α)kGRη( √δn +δ −1/α log δ). With δ = nα/2(α+1) (which grows with n), we get an optimized
log n
generalization bound of C(α)kGRη n1/(2α+2) , with a scaling independent of the dimension d
√ √
(note however that R typically grow with d, i.e., r d, if we have a bound in `∞ -norm for
all our inputs x).
For α = 0, from Prop. 5, the target function belongs
√ to F1 with a norm less than kGRη,
kGRη d
leading to an overall generalization bound of √n .
Note that when the functions fj are exactly the activation functions, the bound is better,
as these functions directly belong to the space F1 .
Multi-dimensional
Pk projection pursuit. We extend the situation above, by assum-
> d×s having all singular values less than
ing f∗ (x) = j=1 Fj (Wj x) with each Wj ∈ R
η and each Fj bounded by ηR and 1-Lipschitz continuous. From Prop. 6, we may ap-
proach each x 7→ Fj (Wj> x) by a function with γ1 -norm less than δηR and uniform ap-
proximation C(α, s)ηRδ −1/(α+(s−1)/2) log δ. This leads to a total approximation error of
kC(α, s)GηRδ −1/(α+(s−1)/2) log δ.
√ √
For α > 1, the estimation error is kGRηδ/ n, with an overall bound of C(α, s)kGRη(δ/ n+
δ −1/(α+(s−1)/2) log δ). With δ = n(α+(s−1)/2)/(2α+s−1) , we get an optimized bound of C(α,s)kGRη
n1/(2α+s+1)
log n.
√
δ√ d
For α = 0, we have an overall bound of C(s)kGRη(δ −2/(s−1) log δ + n
), and with
C(s)kGRη
δ = (n/d)(s−1)/(s+1) , we get a generalization bound scaling as (n/d)1/(s+1)
log(n/d).
27
Bach

function space k · k2 , α > 1 k · k1 , α > 1 α=0

(dq)1/2

>
w x+b d1/2 √ log d 1/2
q( n )
n1/2 n1/2
k
X
fj (wj> x), wj ∈ Rd
kd1/2 kq 1/2 (log d)1/(α+1) k(dq)1/2
1/(2α+2) log n log n

j=1
n n1/(2α+2) n1/2
k kq 1/2 (log d)1/(α+(s+1)/2) (dq)1/2 d1/(s+1)
X
fj (Wj> x), Wj ∈ Rd×s kd1/2
1/(2α+s+1) log n 1/(2α+s+1) log n 1/(s+1) log n

j=1
n n n
Table 2: Summary of generalization bounds with different settings. See text for details.
Note that for s = d and k = 1, we recover the usual Lipschitz-continuous assumption,

with a rate of C(α,d)kGRη
n1/(2α+d+1)
log n.
√ table: when we know a bound r on all dimensions of x, then we may

– Summary √ take
R = r d; this is helpful in comparisons in Table 2, where R is replaced by r d and
the dependence in r is removed as it is the same for all models.
– Dependence on d: when making only a global Lipschitz-continuity assumption, the

generalization bound has a bad scaling in n, i.e., as n−1/(2α+d+1) , which goes down to
zero slowly when d increases. However, when making structural assumptions regarding
the dependence on unknown lower-dimensional subspaces, the scaling in d disappears.
– Comparing different values of α: the value α = 0 always has the best scaling in n,
but constants are better for α > 1 (among which α = 1 has the better scaling in n).
– Bounds for F2 : The simplest upper bound for the penalization by the space F2
depends on the approximation properties√ of F2 . For linear functions and α = 1,
√
it is less than dηR, with a bound GRη √ d . For the other values of α, there is a
n
constant C(d). Otherwise, there is no adaptivity and all other situations only lead to
upper-bounds of O(n−1/(2α+d+1) ). See more details in Section 5.4.
– Sample complexity: Note that the generalization bounds above may be used to obtain
sample complexity results such as dε−2 for affine functions, (εk −1 d−1/2 )−2α−2 for pro-
jection pursuit, and (εk −1 d−1/2 )−s−1−2α for the generalized version (up to logarithmic
terms).
– Relationship to existing work : Maiorov (2006, Theorem 1.1) derives similar results for
neural networks with sigmoidal activation functions (that tend to one at infinity) and
the square loss only, and for a level of smoothness of the target function which grows
with dimension (in this case, once can get easily rates of n−1/2 ). Our result holds
for problems where only bounded first-order derivatives are assumed, but by using
Prop. 2, we would get similar rate by ensuring the target function belongs to F2 and
hence to F1 .
28
Lower bounds. In the sections above, we have only provided generalization bounds.
Although interesting, deriving lower-bounds for the generalization performance when the
target function belongs to certain function classes is out of the scope of this paper. Note
however, that results from Sridharan (2012) suggest that the Rademacher complexities of
the associated function classes provide such lower-bounds. For general Lipschitz-functions,
these Rademacher complexities decreases as n− max{d,2} (von Luxburg and Bousquet, 2004).
5.3 Generalization Bounds for `1 -norm Constraints on Input Weights (p = 1)

We consider the same three situations, assuming that linear predictors have at most q non-
zero elements. We assume that each component of x is almost surely bounded by r (i.e., a
bound in `∞ -norm).
Affine functions. We assume f∗ (x) = w> x + b, with kwk2 6 η and |b| 6 Rη. Given that
√
we have assumed that w has at most q non-zeros, we have kwk1 6 qη.
√
Then, f∗ ∈ F1 with γ1 (f ) 6 C(α)ηr q, with a constant that is independent of d because
we have an affine function. √
Grη q log(d)
From Prop. 7, we thus get a rate of √
n
times a constant (that may depend on α),
which is the same as assuming directly that we optimize over linear predictors only (see, for
example, Bühlmann and Van De Geer, 2011). We recover a high-dimensional phenomenon
√
(although with a slow rate in 1/ n), where d may be much larger than n, as long as log d
√
is small compared to n. The chosen δ is then a constant times rη q (and does not grow
with n).
Projection pursuit. We assume f∗ (x) = kj=1 fj (wj> x), with kwj k2 6 η (which implies
P
√ √
kwj k1 6 qη given our sparsity assumption) and each fj bounded by ηr q and 1-Lipschitz
continuous. We may approach each x 7→ fj (wj> x) by a function with γ1 -norm less than
√ √
δηr q and uniform approximation C(α)ηr qδ −1/α log δ, with a constant that is indepen-
dent of d because we have a function of one-dimensional projection. This leads to a total
√ √
approximation error of kC(α)Gηr qδ −1/α log δ for a norm less than kδηr q.
√
√
For α > 1, the estimation error is kGrηδ√nq log d , with an overall bound of C(α)kGr qη(δ −1/α log δ+
√
√ d)1/(2α+2)
δ √log d
n
). With δ = (n/ log d)α/2(α+1) , we get an optimized bound of C(α)kGr qη log n(log
n1/(2α+2)
,
with a scaling only dependent in d with a logarithmic factor.
√
For α = 0, the target√function belongs to F1 with a norm less than kGr qη, leading to
kGrη√ q log d
an overal bound of n
(the sparsity is not helpful in this case).
Multi-dimensional projection pursuit. We assume f∗ (x) = kj=1 Fj (Wj> x) with each

P
Wj ∈ Rd×s , having all columns with `2 -norm less than η (note that this is a weaker require-
ment than having all singular values that are less than η). If we assume that each of these
√
columns has at most q non-zeros, then the `1 -norms are less than r q and we may use the
approximation properties described at the end of Section 4.6. We also assume that each Fj
√
is bounded by ηr q and 1-Lipschitz continuous (with respect to the `2 -norm).
√
We may approach each x 7→ Fj (Wj> x) by a function with γ1 -norm less than δηr q and
√
uniform approximation C(α, s)ηr qδ −1/(α+(s−1)/2) log δ. This leads to a total approxima-
√
tion error of kC(α, s)Gηr qδ −1/(α+(s−1)/2) log δ.
29
Bach
√ √ √
For α > 1, the estimation error is kGr qηδ√ log d/ n, with an overall bound which is
√
equal to C(α, s)kGr qη(δ −1/(α+(s−1)/2) log δ+ δ √log
n
d
). With δ = (n/ log d)(α+(s−1)/2)/(2α+s−1) ,
√
C(α, s)kGr qη(log d)1/(2α+s+1)
we get an optimized bound of 1/(2α+s+1)
log n.
n√
C(s)kGr qη
For α = 0, we have the bound (n/d)1/(s+1) log(n/d), that is we cannot use the sparsity
as the problem is invariant to the chosen norm on hidden weights.
– High-dimensional variable selection: when k = 1, s = q√and W1 is a projection onto

qη(log d)1/(2α+s+1)
q variables, then we obtain a bound proportional to n1/(2α+s+1)
log n, which
exhibits a high-dimensional scaling in a non-linear setting. Note that beyond sparsity,
no assumption is made (in particular regarding correlations between input variables),
and we obtain a high-dimensional phenomenon where d may be much larger than n.
– Group penalties: in this paper, we only consider `1 -norm on input weights; when
doing joint variable selection for all basis functions, it may be worth using a group
penalty (Yuan and Lin, 2006; Bach, 2008a).
5.4 Relationship to Kernel Methods and Random Sampling

The results presented in the two sections above were using the space F1 , with an L1 -norm on
the outputs weights (and either an `1 - or `2 -norm on input weights). As seen in Sections 2.3
and 3.1, when using an L2 -norm on output weights, we obtain a reproducing kernel Hilbert
space F2 .
As shown in Section 6, the space F2 is significantly smaller than F1 , and in particular is
not adaptive to low-dimensional linear structures, which is the main advantage of the space
F1 . However, algorithms for F2 are significantly more efficient, and there is no need for
the conditional gradient algorithms presented in Section 2.5. The first possibility is to use
the usual RKHS representer theorem with the kernel functions computed in Section 3.1,
leading to a computation complexity of O(n2 ). Alternatively, as shown by Rahimi and Recht
(2007), one may instead sample m basis functions that is m different hidden units, keep the
input weights fixed and optimize only the output layer with a squared `2 -penalty. This will
√
quickly (i.e., the error goes down as 1/ m) approach the non-parametric estimator based
on penalizing by the RKHS norm γ2 . Note that this argument of random sampling has been
used to study approximation bounds for neural networks with finitely many units (Maiorov
and Meir, 2000).
Given the usage of random sampling with L2 -penalties, it is thus tempting to sample
weights, but now optimize an `1 -penalty, in order to get the non-parametric estimator
obtained from penalizing by γ1 . When the number of samples m tends to infinity, we
indeed obtain an approximation that converges to γ1 (this is simply a uniform version of
the law of large numbers). However, the rate of convergence does depend on the dimension
d, and in general exponentially many samples would be needed for a good approximation—
see Bach (2013, Section 6) for a more precise statement in the related context of convex
matrix factorizations.
30
5.5 Sufficient Condition for Polynomial-time Algorithms

In order to preserve the generalization bounds presented above, it is sufficient to be able to
solve the following problem, for any y ∈ Rn and z1 , . . . , zn ∈ Rd+1 :
n
1 X > α

sup yi (v zi )+ , (5)
kvkp =1 n i=1
up to a constant factor. That is, there exists κ > 1, such that for all y and z, we may
compute v̂ such that kv̂kp = 1 and
1 n 1 n
X X
> α
1 > α

n y i (v̂ z i )+ >
κ sup yi (v z i ) + .
kvkp =1 n

i=1 i=1
This is provably NP-hard for α = 0 (see Section 3.2), and for α = 1 (see Section 3.3). If such
an algorithm is available, the approximate conditional gradient presented in Section 2.5 leads
to an estimator with the same generalization bound. Moreover, given the strong hardness
results for improper learning in the situation α = 0 (Klivans and Sherstov, 2006; Livni et al.,
2014), a convex relaxation that would consider a larger set of predictors (e.g., by relaxing
vv > into a symmetric positive-definite matrix), and obtained a constant approximation
guarantee, is also ruled out.
However, this is only a sufficient condition, and a simpler sufficient condition may be
obtained. In the following, we consider V = {v ∈ Rd+1 , kvk2 = 1} and basis functions
ϕv (z) = (v > z)α+ (that is we specialize to the `2 -norm penalty on weight vectors). We
consider a new variation norm γ̂1 which has to satisfy the following assumptions:
– Lower-bound on γ1 : It is defined from functions ϕ̂v̂ , for v̂ ∈ V̂, where for any v ∈ V,
there exists v̂ ∈ V̂ such that ϕv = ϕ̂v̂ . This implies that the corresponding space F̂1
is larger than F1 and that if f ∈ F1 , then γ̂1 (f ) 6 γ1 (f ).
1 n
X

– Polynomial-time algorithm for dual norm: The dual norm sup yi ϕ̂v̂ (zi ) may
v̂∈V̂
n
i=1
be computed in polynomial time.
– Performance guarantees for random direction: There exists κ > 0, such that for any
vectors z1 , . . . , zn ∈ Rd+1 with `2 -norm less than R, and random standard Gaussian
vector y ∈ Rn ,
n
1 X R
sup
yi ϕ̂v̂ (xi ) 6 κ √ . (6)
v̂∈V̂
n n
i=1
We may also replace the standard Gaussian vectors by Rademacher random variables.
We can then penalize by γ̂ instead of γ. Since γ̂1 6 γ1 , approximation properties are

transferred, and because of the result above, the Rademacher complexity for γ̂1 -balls scales
as well as for γ1 -balls. In the next section, we show convex relaxations which cannot achieve
these and leave the existence or non-existence of such norm γ̂1 as an open problem.
31
Bach
6. Convex Relaxations of the Frank-Wolfe Step

In this section, we provide approximation algorithms for the following problem of maximiz-
ing, for a given y ∈ Rn and vectors z1 , . . . , zn :
n
1X
sup yi (v > zi )α+
kvkp =1 n
i=1
These approximation algorithms may be divided in three families, as they may be based
on (a) geometric interpretations as linear binary classification or computing Haussdorff
distances (see Section 3.2 and Section 3.3), (b) on direct relaxations, on (c) relaxations
of sign vectors. For simplicity, we only focus on the case p = 2 (that is `2 -constraint on
weights) and on α = 1 (rectified linear units). As described in Section 5.5, constant-factor
approximation ratios are not possible, while approximation ratios that increases with n
are possible (but as of now, we only obtain scalings in n that provide a provable sample
complexity with a polynomial time algorithm which is exponential in the dimension d.
6.1 Semi-definite Programming Relaxations

We present two relaxations, which are of the form described in Section 5.5 (leading to
potential generalization bounds) but do not attain the proper approximation scaling (as
was checked empirically).
Note that all relaxations that end up being Lipschitz-continuous functions of z, will have
at least the same scaling than the set of these functions.
q The Rademacher complexity of
√
such functions is well-known, that is 1/ n for d = 1, logn n for d = 2 and n−1/d for larger
d (von Luxburg and Bousquet, 2004). Unfortunately, the decay in n is too slow to preserve
√
generalization bounds (which would require a scaling in 1/ n).
d-dimensional relaxation. We denote ui = (v > zi )+ = 21 v > zi + 12 |v > zi |. We may then
q
> > > >
use 2ui − v zi = |v zi | and, for kvk2 = 1, kvv zi k2 = |v zi | = zi> vv > zi . By denoting
V = vv > , the constraint that ui = (v > zi )+ = 12 v > zi + 12 |v > zi | is equivalent to
q
>
kV zi k2 6 2ui − v zi 6 zi> V zi and V < 0, tr V = 1, rank(V ) = 1.
We obtain a convex relaxation when removing the rank constraint, that is

q
sup u> y such that ∀i ∈ {1, . . . , n}, kV zi k2 6 2ui − v > zi 6 zi> V zi .
V <0, tr V =1, u∈Rn
(n + d)-dimensional relaxation. We may go further by also considering quadratic forms

in u ∈ Rn defined above. Indeed, we have:
(2ui − v > zi )(2uj − v > zj ) = |v > zi | · |v > zj | = |v > zi zj> v| = | tr V zi zj> |,
which leads to a convex program in U = uu> , V = vv > and J = uv > , that is a semidefinite
program with d + n dimensions, with the constraints
4Uij + x> > > >

j V zi − 2δi Jzj − 2δj Jzi > | tr V zi zj |,
32
U J u u >
and the usual semi-definite contraints < , with the additional
J> V v v
constraint that 4Uii + zi> V zi − 4δi> Jzi = tr V zi zi> .
If we add these constraints on top of the ones above, we obtain a tighter relaxation. Note
that for this relaxation, we must have [(2ui −v > zi )−(2uj −v > zj )] less than a constant times
kzi − zj k2 . Hence, the result mentioned above regarding Lipschitz-continuous functions and
the scaling of the upper-bound for random y holds (with the dependence on n which is not
good enough to preserve the generalization bounds with a polynomial-time algorithm).
6.2 Relaxation of Sign Vectors

By introducing a sign vector s ∈ Rn such that si ∈ {−1, 1} and si v > xi = |v > xi |, we have
the following relaxation with S = ss> , V = vv > and J = sv > :
S J s s >
– Usual semi-definite constraint: < ,
J> V v v
– Unit/trace constraints: diag(S) = 1 and tr V = 1,
– Sign constraint: δi> Jxi > maxj6=i |δj> Jxi |.
– Additional constraint: (x>

i V xi )
1/2 6 δ > Jx .
i i
1 Pn > 1 Pn >
We then need to maximize 2n i=1 yi δi Jxi + 2n i=1 yi v xi , which leads to a semidefinte
program. Again empirically, it did not lead to the correct scaling as a function of n for
random Gaussian vectors y ∈ Rn .
7. Conclusion
In this paper, we have provided a detailed analysis of the generalization properties of convex
neural networks with positively homogenous non-decreasing activation functions. Our main
new result is the adaptivity of the method to underlying linear structures such as the
dependence on a low-dimensional subspace, a setting which includes non-linear variable
selection in presence of potentially many input variables.
All our current results apply to estimators for which no polynomial-time algorithm is
known to exist and we have proposed sufficient conditions under which convex relaxations
could lead to the same bounds, leaving open the existence or non-existence of such algo-
rithms. Interestingly, these problems have simple geometric interpretations, either as binary
linear classification, or computing the Haussdorff distance between two zonotopes.
In this work, we have considered a single real-valued output; the functional analysis
framework readily extends to outputs in a finite-dimensional vector-space where vector-
valued measures could be used, and then apply to multi-task or multi-class problems. How-
ever, the extension to multiple hidden layers does not appear straightforward as the units
of the last hidden layers share the weights of the first hidden layers, which should require a
new functional analysis framework.
33
Bach
Acknowledgements
The author was partially supported by the European Research Council (SIERRA Project),
and thanks Nicolas Le Roux for helpful discussions. The author also thanks Varun Kanade
for pointing the NP-hardness linear classification result. The main part of this work was
carried through while visiting the Centre de Recerca Matemàtica (CRM) in Barcelona.
Appendix A. Reproducing Kernel Hilbert Spaces for `2 -norm

Penalization
In this section, we consider a Borel probability measure τ on the compact space V, and
functions ϕv : X → R such that the functions v 7→ ϕv (x) are measurable for all x ∈ X .
We study the set F2 ofR functions f such that there exists a squared-integrable function
p : X → R with R f (x) = V p(v)ϕv (x)dτ (v) for all x ∈ X . For f ∈ F2 , we define γ22 (f ) as the
2
infimum of V p(v) dτR(v) over all decompositions of f . We now show that F2 is an RKHS
with kernel k(x, y) = V ϕv (x)ϕv (y)dτ (v).
We follow the proof of Berlinet and Thomas-Agnan (2004, Section 4.1) and extend it to
integrals ratherRthan finite sums. We consider the linear mapping T : L2 (dτ ) → F2 defined
by (T p)(x) = V p(v)ϕv (x)dτ (v), with null space K. When restricted to the orthogonal
complement KR⊥ , we obtain a bijection U from K⊥ to F2 . We then define a dot-product on
F2 as hf, gi = V (U −1 f )(v)(U −1 g)(v)dτ (v).
We first show that this defines an RKHS with kernel k defined above. For this, we triv-
ially have k(·, y) ∈ F2 for all y ∈ X . Moreover, for any y ∈ X , we have R with p = U −1 k(·, y) ∈
⊥ −1
RK and −1
q : v 7→ ϕv (y), p −Rq ∈ K, which implies that hf, k(·, y)i = V (U f )(v)p(v)dτ (v) =
V (U f )(v)q(v)dτ (v) = V (U −1 f )(v)ϕv (y)dτ (v) = T (U −1 f )(y) = f (y), hence the repro-
ducing property is satisfied. Thus, F2 is an RKHS.
We now need to show that the RKHS norm which we have defined is actually γ2 . For
R f ∈2 F2 such that
any f = T p, for p ∈ L2 (dτ ), we have p = U −1 f + q, where q ∈ K. Thus,
p(v) dτ (v) = kpkL2 (dτ ) = kU −1 f k2L2 (dτ ) + kqk2L2 (dτ ) = kf k2 + kqk2L2 (dτ ) . This implies that
2
RV 2 2
V p(v) dτ (v) > kf k with equality if and only if q = 0. This shows that γ2 (f ) = kf k.
Appendix B. Approximate Conditional Gradient with Multiplicative

Oracle
In this section, we wish to minimize a smooth convex functional J(h) on for h in a Hilbert-
space over a norm ball {γ(h) 6 δ}. A multiplicative approximate oracle outputs for any
g ∈ Rn , ĥ such that γ(ĥ) = 1, and
hĥ, gi 6 max hh, gi 6 κ hĥ, gi,

γ(h)61
for a fixed κ > 1. We now propose a modification of the conditional gradient algorithm
that converges to a certain h such that γ(h) 6 δ and for which inf γ(h)6δ J(h) 6 J(ĥ) 6
inf γ(h)6δ/κ J(h).
34
We assume the smoothness of the function J with respect to the norm γ, that is, for a
certain L > 0, for all h, h0 such that γ(h) 6 δ, then
L
J(h0 ) 6 J(h) + hJ 0 (h), h0 − hi + γ(h − h0 )2 . (7)
2
We consider the following recursion
ĥt = −δ × output of approximate oracle at −J 0 (ht )

ht+1 ∈ arg min J((1 − ρ)ht + ρĥt ).
ρ∈[0,1]
In the previous recursion, one may replace the minimization of J on the segment [ht , ĥt ]
with the minimization of its upper-bound of Eq. (7) taken at h = ht . From the recursion,
all iterates are in the γ-ball of radius δ. Following the traditional convergence proof for the
conditional gradient method (Dunn and Harshbarger, 1978; Jaggi, 2013), we have, for any
ρ in [0, 1]:
J(ht+1 ) 6 J(ht ) − ρhJ 0 (ht ), ht − ĥt i + 2Lρ2 δ 2

ĥt
= J(ht ) − ρJ 0 (ht )> ht + κhJ 0 (ht ), i + 2Lρ2 δ 2
κ
6 J(ht ) − ρJ 0 (ht )> ht − max { − hJ 0 (ht ), hi} + 2Lρ2 δ 2 .
γ(h)6δ/κ
If we take h∗ the minimizer of J on {γ(h) 6 δ/κ}, we get:
J(ht+1 ) 6 J(ht ) − ρhJ 0 (ht ), ht − h∗ i + 2Lρ2 δ 2 .
Then, by using J(ht ) > J(h∗ ) + hJ 0 (ht ), h∗ − ht i, we get:
J(ht+1 ) − J(h∗ ) 6 (1 − ρ)[J(ht ) − J(h∗ )] + 2Lρ2 δ 2 .
This is valid for any ρ ∈ [0, 1]. If J(ht ) − J(h∗ ) 6 0 for some t, then by taking ρ = 0
it remains the same of all greater t. Therefore, up to (the potentially never happening)
point where J(ht ) − J(h∗ ) 6 0, we can apply the regular proof of the conditional gradien
2 2
to obtain: J(ht ) 6 inf γ(h)6δ/κ J(h) + 4Lρt δ , which leads to the desired result. Note that a
similar reasoning may be used for ρ = 2/(t + 1).
Appendix C. Proofs for the 2-dimensional Sphere (d = 1)

In this section, we consider only the case d = 1, where the sphere Sd is isomorphic to [0, 2π]
(with periodic boundary conditions). We Rmay then compute the norm γ2 in closed form.
1 2π
Indeed, if we can decompose g as g(θ) = 2π 0 p(ϕ)σ(cos(ϕ−θ))dϕ, then the decomposition
of g into the k-th frequency elements (the combination of the two k-th elements of the Fourier
series) is equal to, for σ(u) = (u)α+ , and for k > 0:
Z 2π
1
gk (θ) = g(η) cos k(θ − η)dη
π 0
35
Bach
1 2π 1
Z Z 2π
= p(ϕ)σ(cos(η − ϕ))dϕ cos k(θ − η)dη through the decomposition of g,
π 0 2π 0
Z 2π Z 2π
1
= p(ϕ) σ(cos(η − ϕ)) cos k(θ − η)dη dϕ
2π 2 0 0
Z 2π Z 2π
1
= p(ϕ) σ(cos η) cos k(θ − ϕ − η)dη dϕ by a change of variable,
2π 2 0 0
Z 2π Z 2π
1
= p(ϕ) cos k(θ − ϕ) σ(cos η) cos kη dη
2π 2 0 0
Z 2π
+ sin k(θ − ϕ) σ(cos η) sin kη dη dϕ by expanding the cosine,
0
Z 2π Z 2π
1 1
= σ(cos η) cos kη dη p(ϕ) cos k(θ − ϕ) + 0 by a parity argument,
2π 0 π 0
Z 2π
1
= λk pk (θ) with λk = σ(cos η) cos kη dη.
2π 0
For k = 0, the same equality holds (except that the two coefficients g0 and p0 are divided
by 2π except of π).
Thus we may express kpk2L2 (Sd ) as
X X X
kpk2L2 (Sd ) = kpk k2L2 (Sd ) = kpk k2L2 (Sd ) + kpk k2L2 (Sd )
k>0 λk 6=0 λk =0
X 1 X
= 2 kgk k2L2 (Sd ) + kpk k2L2 (Sd ) .
λk
λk 6=0 λk =0
If we minimize over p, we thus need to have kpk k2L2 (Sd ) = 0 for λk = 0, and we get
X 1
γ2 (g)2 = kgk k2L2 (Sd ) . (8)
λ2k
λk 6=0
We thus simply need to compute λk and its decay for all values of α, and then relate them
to the smoothness properties of g, which is standard for Fourier series.
C.1 Computing λk
1
R 2π
We now detail the computation of λk = 2π 0 σ(cos η) cos kη dη for the different functions
σ = (·)α+ . We have for α = 0:
Z 2π Z π/2
1 1 1 kπ
1cos η>0 cos kη dη = cos kη dη = sin if k 6= 0.
2π 0 2π −π/2 πk 2
For k = 0 it is equal to 21 . It is equal to zero for all other even k, and different from zero
for all odd k, with λk going to zero as 1/k.
We have for α = 1:
Z 2π Z π/2
1 1
(cos η)+ cos kη dη = cos η cos kη dη
2π 0 2π −π/2
36
Z π/2
1 1 1
= [ cos(k + 1)η + cos(k − 1)η] dη
2π −π/2 2 2

1 2 π 2 π
= sin(k + 1) + sin(k − 1)
4π k + 1 2 k−1 2
cos kπ − cos kπ

2 1 1 2
= − = for k 6= 1.
2π k+1 k−1 π(k 2 − 1)
For k = 1, it is equal to 1/4. It is equal to zero for all other odd k, and different from zero
for all even k, with λk going to zero as 1/k 2 .
For α = 2, we have:
Z 2π Z π/2 Z π/2
1 1 1 1 + cos 2η
(cos η)2+ cos kη dη = (cos η)2 cos kη dη = cos kη dη
2π 0 2π −π/2 2π −π/2 2
Z π/2
1 1 1 1
= [ cos kη + cos(k + 2)η + cos(k − 2)η] dη
2π −π/2 2 4 4

1 2 π 1 π 1 π
= sin k + sin(k + 2) + sin(k − 2)
4π k 2 k+2 2 k−2 2
sin(k π2 ) 2

1 1
= − −
4π k k+2 k−2
sin(k π2 ) 2k 2 − 8 − k 2 + 2k − k 2 − 2k

=
4π k(k 2 − 4)
−8 sin(k π2 )
= for k ∈
/ {0, 2}.
4πk(k 2 − 4)
For k = 0, it is equal to 1/4, and for k = 2, it is equal to 1/8. It is equal to zero for all
other even k, and different from zero for all odd k, with λk going to zero as 1/k 3 .
The general case for α > 2 will be shown for for all d in Appendix D.2: for all α ∈ N,
λk is different from zero for k having the opposite parity of α, with a decay as 1/k α+1 . All
values from k = 0 to α are also different from zero. All larger values with the same parity
as α are equal to zero.
C.2 Proof of Prop. 2 for d = 1

We only consider the proof for d = 1. For the proof for general d, see Appendix D.3.
Given the zero values of λk given above, if g has the opposite parity than α (that is, is
even when α is odd, and vice-versa), then we may define p through its Fourier series, which
is obtained by multiplying the one of g by a strictly positive sequence growing as k α+1 .
Thus, if g is such that its (α+1)-th order derivative is squared-integrable, then p defined
above is squared-integrable, that is, g ∈ G2 . Moreover, if all derivatives of order less than
(α + 1) are bounded by η, p is squared-integrable and kpk2L2 (Sd ) is upper-bounded by a
constant times η 2 , i.e., γ2 (g)2 6 C(α)2 η 2 .
Note that we could relax the assumption that g is even (resp. odd) by adding all trigono-
metric polynomials of order less than α.
37
Bach
C.3 Proof of Prop. 3 for d = 1

Again, we only consider the proof for d = 1. For the proof for general d, see Appendix D.4.
Without loss of generality, we assume that η = 1. For d = 1, we essentially want to
approximate a Lipschitz-continuous function by a function which is (α + 1)-times differen-
tiable.
For α = 0, then the function g is already in G2 with a norm less than one, because
Lipschitz-continuous functions are almost everywhere differentiable with bounded deriva-
tive (Adams and Fournier, 2003). We thus now consider α > 0.
Given λk defined above and r ∈ (0, 1), we define p̂ through
X
p̂k (θ) = λ−1 k
k gk (θ)r .
k,λk 6=0
Our goal is to show that for r chosen close enough to 1, then the function ĝ defined from p̂
has small enough norm γ2 (ĝ) 6 kp̂kL2 (Sd ) , and is close to g.
Computation of norm. We have
X
kp̂k2L2 (Sd ) = λ−2 2k 2
k r kgk kL2 (Sd ) .
k,λk 6=0
Since g is 1-Lipschitz-continuous with constant 1, then it has a squared-integrable derivative

f = g 0 with norm less than 1 (Adams and Fournier, 2003), so that
X
kf k2L2 (Sd ) = kfk k2L2 (Sd ) 6 1.
k>0
This implies that using λ−1

k = O(k
α+1 ):
kp̂k2L2 (Sd ) 6 λ−2 2 0 2

0 kg0 kL2 (Sd ) + kg kL2 (Sd ) max r2k λ−2
k k
−2
6 C + Ckg 0 k2L2 (Sd ) max r2k k 2α ,
k>1,λk 6=0 k>1
because kg0 k2L2 (Sd ) and kf k2L2 (Sd ) are bounded by 1.

We may now compute the derivative of k 7→ r2k k 2α with respect to k (now considered a
real number), that is 2αk 2α−1 r2k + k 2α r2k 2 log r, which is equal to zero for αk = log 1r , that
is k = logα 1 , the maximum being then e−2α ( logα 1 )2α = O((1 − r)−2α ), by using the concavity
r r
of the logarithm. Thus kp̂kL2 (Sd ) 6 C(1 − r)−α . This defines ĝ with γ(ĝ) 6 C(1 − r)−α .
Computing distance between ĝ and g. We have:
X X 1 Z 2π 1
Z 2π
k k
ĝ(θ) = gk (θ)r = g(η)r cos k(θ − η)dη + g(η)dη
π 0 2π 0
k>0 k>0
1 2π X k
Z Z 2π
1
= r cos k(θ − η) g(η)dη − g(η)dη
π 0 2π 0
k>0
1 2π
Z Z 2π
1 1
= Real g(η)dη − g(η)dη
π 0 1 − rei(θ−η) 2π 0
38
1 2π
Z 2π
1 − r cos(θ − η)
Z
1
= g(η)dη − g(η)dη
π 0 (1 − r cos(θ − η))2 + r2 (sin(θ − η))2 2π 0
1 2π
Z 2π
1 − r cos(θ − η)
Z
1
π 0 1 + r2 − 2r cos(θ − η) 2π 0
Z 2π Z 2π
1 − r2 + 1 + r2 − 2r cos(θ − η)

1 1
2π 0 1 + r2 − 2r cos(θ − η) 2π 0
Z 2π
1 − r2

1
= g(η)dη.
2π 0 1 + r2 − 2r cos(θ − η)
We have, for any θ ∈ [0, 2π]
Z 2π
1 − r2

1
|ĝ(θ) − g(θ)| = [g(η) − g(θ)]dη
2π 0 1 + r2 − 2r cos(θ − η)
Z 2π
1 − r2

1
6 |g(η) − g(θ)|dη
2π 0 1 + r2 − 2r cos(θ − η)
Z π
1 − r2

1
= |g(θ) − g(θ + η)|dη by periodicity,
2π π 1 + r2 − 2r cos η
1 π/2 1 − r2
Z
= |g(θ) − g(θ + η)|dη by parity of g,
π π/2 1 + r2 − 2r cos η
1 π/2 1 − r2 √
Z
6 2
2| sin η|dη
π π/2 1 + r − 2r cos η
because the distance on the sphere is bounded by the sine,
Z π
1 − r2

2
6 sin η dη
π 0 1 + r2 − 2r cos η
1 1 1 − r2
Z
= dt by the change of variable t = cos θ,
π 0 1 + r2 − 2rt
Z 1
1
6 C(1 − r) dt
0 1 + r2 − 2rt
1
−1 1 + r2

2 1
= C(1 − r) log(1 + r − 2rt) = C(1 − r) log .
2r 0 2r (1 − r)2
It can be easily checked that for any r ∈ (1/2, 1), the last function is less than a constant
times 52 C(1 − r) log 1−r
1
. We thus get for δ large enough, by taking r = 1 − (C/δ)1/α ∈
(1/2, 1), an error of
(C/δ)1/α log(C/δ)−1/α = O(δ −1/α log δ).
This leads to the desired result.
Appendix D. Approximations on the d-dimensional Sphere

In this appendix, we first review tools from spherical harmonic analysis, before proving
the two main propositions regarding the approximation properties of the Hilbert space G2 .
Using spherical harmonics in our set-up is natural and is common in the analysis of ridge
functions (Petrushev, 1998) and zonotopes (Bourgain and Lindenstrauss, 1988).
39
Bach
D.1 Review of Spherical Harmonics Theory

In this section, we review relevant concepts from spherical harmonics. See Frye and
Efthimiou (2012); Atkinson and Han (2012) for more details. Spherical harmonics may
be seen as extension of Fourier series to spheres in dimensions more than 2 (i.e., with our
convention d > 1).
For d > 1, we consider the sphere Sd = {x ∈ Rd+1 , kxk2 = 1} ⊂ Rd+1 , as well as its
2π (d+1)/2
normalized rotation-invariant measure τd (with mass 1). We denote by ωd = Γ((d+1)/2) the
surface area of the sphere Sd .
Definition and links with Laplace-Beltrami operator. For any k > 1 (for k = 0,
the constant function is the corresponding basis element), there is an orthonormal basis
d 2k+d−1 k+d−2

of spherical harmonics,
R Ykj : S → R, 1 6 j 6 N (d, k) = k d−1 . They are such
hYki , Ysi iSd = Sd Yki (x)Ysj dτd (x) = δij δsk .
Each of these harmonics may be obtained from homogeneous polynomials in Rd with an
Euclidean Laplacian equal to zero, that is, if we define a function Hk (y) = Yki (y/kyk2 )kykk2
for y ∈ Rd+1 , then Hk is a homogeneous polynomial of degree k with zero Laplacian. From
the relationship between the Laplacian in Rd+1 and the Laplace-Beltrami operator ∆ on Sd ,
Yki is an eigenfunction of ∆ with eigenvalue −k(k + d − 1). Like in Euclidean spaces, the
Laplace-Beltrami operator may be used to characterize differentiability of functions defined
on the sphere (Frye and Efthimiou, 2012; Atkinson and Han, 2012).
Legendre polynomials. We have the addition formula
N (d,k)
X
Ykj (x)Ykj (y) = N (d, k)Pk (x> y),
j=1
where Pk is a Legendre polynomial of degree k and dimension d + 1, defined as (Rodrigues’

formula):
Γ(d/2) d k
Pk (t) = (−1/2)k (1 − t2 )(2−d)/2 (1 − t2 )k+(d−2)/2 .
Γ(k + d/2) dt
They are also referred to as Gegenbauer polynomials. For d = 1, Pk is the k-th Chebyshev
polynomial, such that Pk (cos θ) = cos(kθ) for all θ (and we thus recover the Fourier series
framework of Appendix C). For d = 2, Pk is the usual Legendre polynomial.
The polynomial Pk is even (resp. odd) when k is even (resp. odd), and we have
1
ωd
Z
1
Pk (t)Pj (k)(1 − t2 )(d−2)/2 dt = δjk .
−1 ωd−1 N (d, k)
−1 2
For small k, we have P0 (t) = 1, P1 (t) = t, and P2 (t) = (d+1)t
d .
The Hecke-Funk formula leads to, for any linear combination Yk of Ykj , j ∈ {1, . . . , N (d, k)}:
1
ωd−1
Z Z
>
f (x y)Yk (y)dτd (y) = Yk (x) f (t)Pk (t)(1 − t2 )(d−2)/2 dt.
Sd ωd −1
40
Decomposition
R of functions in L2 (Sd ). Any function g : Sd → R, such that we have
Sd g(x)dτd (x) = 0 may be decomposed as
∞ NX
X (d,k) ∞ NX
X (d,k) Z
g(x) = hYkj , giYkj (x) = Ykj (y)Ykj (x)g(y)dτd (y)
k=1 j=1 k=1 j=1 Sd
X∞ Z
= gk (x) with gk (x) = N (d, k) g(y)Pk (x> y)dτd (y).
k=1 Sd
This is the decomposition in harmonics of degree k. Note that

Z
>
g1 (x) = x d yg(y)dτd (y)
Sd
is the linear part of g (i.e., >

R if g(x) = w x, g1 = g). Moreover, if g does not have zero mean,
we may define g0 (x) = Sd g(y)dτd (y) as the average value of g. Since the harmonics of
different degrees are orthogonal to each other, we have the Parseval formula:
X
kgk2Sd = kgk k2Sd .
k>0
Decomposition of functions of one-dimensional projections. If g(x) = ϕ(x> v) for

v ∈ Sd and ϕ : [−1, 1] → R, then
Z
gk (x) = N (d, k) ϕ(v > y)Pk (x> y)dτ (y)
S d
Z 1
ωd−1 >
= N (d, k) Pk (x v) ϕ(t)Pk (t)(1 − t2 )(d−2)/2 dt
ωd −1
Z 1 NX(d,k)
ωd−1 > 2 (d−2)/2
= Pk (x v) ϕ(t)Pk (t)(1 − t ) dt Ykj (x)Ykj (y),
ωd −1 j=1
and
1 2 NX
(d,k)
ωd−1
Z
kgk k2L2 (Sd ) = Pk (x> v) ϕ(t)Pk (t)(1 − t )2 (d−2)/2
dt Ykj (y)2
ωd −1 j=1
1 2
ωd−1
Z
= Pk (x> v) 2 (d−2)/2
ϕ(t)Pk (t)(1 − t ) dt N (d, k)Pk (1)
ωd −1
1 2
ωd−1
Z
= Pk (x> v) ϕ(t)Pk (t)(1 − t2 )(d−2)/2 dt N (d, k).
ωd −1
D.2 Computing the RKHS Norm γ2

Like for the case d = 1, we may compute the RKHS norm γ2 of a function
P g in closed
form given its decomposition in the basis of spherical harmonics g = k>0 gk . If we can
41
Bach
p(w)σ(w> x)dτd (w) for a certain function p : Sd → R, then we have,

R
decompose g(x) = Sd
for k > 0:
Z
gk (x) = N (d, k) g(y)Pk (x> y)dτd (y)
Z Z Sd
= N (d, k) p(w)σ(w> y)Pk (x> y)dτd (y)dτd (w)

d d
ZS S Z
> >
= N (d, k) p(w) σ(w y)Pk (x y)dτd (y) dτd (w)
Sd Sd
Z 1
ωd−1
Z
= N (d, k) p(w)Pk (x> w) σ(t)Pk (t)(1 − t2 )(d−2)/2 dt dτd (w)
ωd Sd −1
using the Hecke-Funk formula,
Z 1
ωd−1
= λk pk (x) with λk = σ(t)Pk (t)(1 − t2 )(d−2)/2 dt.
ωd −1
R1
If k ≡ α mod. 2, then λk ∝ 21 −1 tα Pk (t)(1 − t2 )(d−2)/2 dt = 0, for k > α since Pk is
orthogonal to all polynomials of degree strictly less than k for that dot-product. Other-
wise, λk 6= 0, since tα may be decomposed as combination with non-zero coefficients of
polynomials Pj for j ≡ α mod. 2, j 6 α.
We now provide an explicit formula extending the proof technique (for α = 1) of Schnei-
der (1967) and Bourgain and Lindenstrauss (1988) to all values of α. See also Mhaskar
(2006).
We have, by α successive integration by parts, for k > α + 1:
Z 1
d k
tα (1 − t2 )k+(d−2)/2 dt
0 dt
Z 1
α d k−α 2 k+(d−2)/2 α
d k−α−1
2 k+(d−2)/2

= (−1) α! (1 − t ) dt = −(−1) α! (1 − t )
0 dt dt
t=0
d k−α−1 X k + (d − 2)/2

= −(−1)α α! (−1)j t2j

using the binomial formula,
dt j t=0
j>0
d k−α−1 k + (d − 2)/2
α j 2j

= −(−1) α! (−1) t for 2j = k − α − 1,
dt j t=0
k + (d − 2)/2

α
= −(−1) α! (−1)j (2j)! for 2j = k − α − 1.
j
Thus
k + (d − 2)/2

ωd−1 Γ(d/2)
λk = − (−1/2)k (−1)α α! (−1)j (2j)! for 2j = k − α − 1,
ωd Γ(k + d/2) j
ωd−1 Γ(d/2) Γ(k + d2 )
= − (−1/2)k (−1)α α! d
(−1)j Γ(2j + 1)
ωd Γ(k + d/2) Γ(j + 1)Γ(k + 2 − j)
ωd−1 Γ(d/2) Γ(k + d )(−1)(k−α−1)/2 Γ(k − α)
= − (−1/2)k (−1)α α! k 2α 1
ωd Γ(k + d/2) Γ( 2 − 2 + 2 )Γ( k2 + d2 + α2 + 21 )
d − 1 α!(−1)(k−α−1)/2 Γ(d/2)Γ(k − α)
= .
2π 2k Γ( 2 − 2 + 21 )Γ( k2 + d2 +
k α α
2 + 21 )
42
√
By using Stirling formula Γ(x) ≈ xx−1/2 e−x 2π, we get an equivalent when k or d tends
to infinity as a constant (that depends on α) times
dd/2+1/2 k k/2−α/2+1/2 (k + d)−k/2−d/2−α/2 .
Note that all exponential terms cancel out. Moreover, when k tends to infinity and d is
considered constant, then we get the equivalent k −d/2−α−1/2 , which we need for the following
sections. Finally, when d tends to infinity and k is considered constant, then we get the
equivalent d−α/2−k/2+1/2 .
We will also need expressions of λk for k = 0 and k = 1. For k = 0, we have:
1 1 √
du
Z Z
α 2 d/2−1
t (1 − t ) dt = (1 − u)α/2 ud/2−1 √ with t = 1 − u,
0 0 2 1−u
Z 1
1 1 Γ(α/2 + 1/2)Γ(d/2)
= (1 − u)α/2+1/2−1 ud/2−1 du = ,
2 0 2 Γ(α/2 + 1/2 + d/2)
using the normalization factor of the Beta distribution. This leads to
ωd−1 1 Γ(α/2 + 1/2)Γ(d/2) d − 1 1 Γ(α/2 + 1/2)Γ(d/2)

λ0 = = ,
ωd 2 Γ(α/2 + 1/2 + d/2) 2π 2 Γ(α/2 + 1/2 + d/2)
which is equivalent to d1/2−α/2 as d tends to infinity.

Moreover, for k = 1, we have (for α > 0):
1 d 1 1
du
Z Z Z
tα (1 − t2 )d/2 dt = −α tα−1 (1 − t2 )d/2 dt = −α
(1 − u)α/2−1/2 ud/2 √
0 dt 0 0 2 1−u
Z 1
Γ(α/2)Γ(d/2 + 1)
= −α/2 (1 − u)α/2−1 ud/2+1−1 du = −α/2 .
0 Γ(α/2 + d/2 + 1)
This leads to, for α > 0:
2d−1 Γ(α/2)Γ(d/2 + 1) d − 1 α Γ(α/2)Γ(d/2 + 1)

λ1 = (−1/2) (−α/2) = ,
d 2π Γ(α/2 + d/2 + 1) d 4π Γ(α/2 + d/2 + 1)
which is equivalent to d−α/2 as d tends to infinity.

Finally, for α = 0, λ1 = d−1
2dπ . More generally, we have |λk | ∼ C(d)k
−(d−1)/2−α−1 .
Computing the RKHS norm. Given g with the correct parity, then we have
X
γ2 (g)2 = λ−2 2
k kgk kL2 (Sd ) .
k>0
D.3 Proof of Prop. 2 for d > 1

Given the expression of λk from the section above, the proof is essentially the same than for
d = 1 in Appendix C.3. If g is s-times differentiable with all derivatives bounded uniformly
by η, then is equal to g = ∆s/2 f for a certain function f such that kf kL2 (Sd ) 6 η (where ∆
is the Laplacian on the sphere) (Frye and Efthimiou, 2012; Atkinson and Han, 2012).
43
Bach
Moreover, since g has the correct parity,

X
γ2 (g)2 6 kpk2L2 (Sd ) 6 λ−2 2
k kgk kL2 (Sd )
k>1,λk 6=0
Also, gk are eigenfunctions of the Laplacian with eigenvalues k(k + d − 1). Thus, we have
1
kgk k22 6 kfk k2L2 (Sd ) 6 kfk k2L2 (Sd ) /k 2s ,
[k(k + d − 1)]s
leading to γ2 (g)2 6 max λ−2

k k
−2s
kf k2L2 (Sd ) 6 max k d−1+2α+2 k −2s kf k2L2 (Sd ) 6 C(d)η 2 , if
k>2 k>2
s > (d − 1)/2 + α + 1, which is the desired result.
D.4 Proof of Prop. 3 for d > 1

Without loss of generality we assume that η = 1, and we follow the same proof as for d = 1
in√Appendix
p C.3. We have assumed that for all x, y ∈ Sd , |g(x) − g(y)| 6 ηkx − yk2 =
η 2 1 − x> y. Given the decomposition in the k-th harmonics, with
Z
gk (x) = N (d, k) g(y)Pk (x> y)dτd (y),
Sd
we may now define, for r < 1:

X
p̂(x) = λ−1 k
k r gk (x),
k,λk 6=0
which is always defined when r ∈ (0, 1) because the series is absolutely convergent. This
defines a function ĝ that will have a finite γ2 -norm and be close to g.
Computing the norm. Given our assumption regarding the Lipschitz-continuity of g,
we have g = ∆1/2 f with f ∈ L2 (Sd ) with norm less than 1 (Atkinson and Han, 2012).
Moreover kgk k2L2 (Sd ) 6 Ck 2 kfk k2L2 (Sd ) . We have
X
kp̂k2L2 (Sd ) = λ−2 2k 2
k r kgk kL2 (Sd )
k,λk 6=0
6 C(d, α) max k d−1+2α r2k kf k2L2 (Sd ) because λk = Ω(k −d/2−α−1/2 ),

k>0
6 C(d, α)(1 − r)−d+1−2α (see Appendix C.3).
The function p̂ thus defines a function ĝ ∈ G1 by ĝk = λk pk , for which γ2 (g) 6 C(d, α)(1−
r)(−d+1)/2−α .
Approximation properties. We now show that g and ĝ are close to each other. Because
of the parity of g, we have ĝk = rk gk . We have, using Theorem 4.28 from Frye and Efthimiou
(2012):
X X Z
ĝ(x) = k
r = k
r N (d, k) g(y)Pk (x> y)dτd (y)
k>0 k>0 Sd
44
Z X
k >
= g(y) r N (d, k)Pk (x y) dτd (y)
Sd k>0
1 − r2
Z
= g(y) dτd (y).
Sd (1 + r2 − 2r(x> y))(d+1)/2
Moreover, following Bourgain and Lindenstrauss (1988), we have:
1 − r2
Z
g(x) − ĝ(x) = [g(x) − g(y)] dτd (y)
Sd (1 + r2 − 2r(x> w))(d+1)/2
1 − r2
Z
= 2 [g(x) − g(y)] dτd (y) by parity of g,
Sd , y > x>0 (1 + r2 − 2r(x> w))(d+1)/2
Z √ p 1 − r2
|g(x) − ĝ(x)| 6 2 1 − x> y dτd (y).
Sd , y > x>0 (1 + r2 − 2r(x> y))(d+1)/2
As shown by Bourgain and Lindenstrauss (1988, Eq. (2.13)), this is less than a constant
1
that depends on d times (1 − r) log 1−r . We thus get for δ large enough, by taking 1 − r =
(C/δ) 1/(α+(d−1)/2) ∈ (0, 1), an error of
(C/δ)1/(α+(d−1)/2) log(C/δ)−1/(α+(d−1)/2) ] = O(δ 1/(α+(d−1)/2) log δ),
which leads to the desired result.
D.5 Finding Differentiable Functions which are not in G2

In this section, we consider functions on the sphere which have the proper parity with
respect to α, which are s-times differentiable with bounded derivatives, but which are not
in G2 . We then provide optimal approximation rates for these functions.
We assume that s − α is even, we consider g(x) = (w> x)s+ for d
a certain arbitrary w ∈ S .
ωd−1 >
R1
As computed at the end of Appendix D.1, we have kgk k2L2 (Sd ) = ωd Pk (x v) −1 ϕ(t)Pk (t)(1−
2
2
t )(d−2)/2 dt N (d, k). Given the computations from Appendix D.2, the squared norm equal
2
ωd−1 > v) 1 ϕ(t)P (t)(1 − t2 )(d−2)/2 dt goes down to zero as k −d−2s−1 , while
R
to ωd P k (x −1 k
N (d, k) grows as k d−1 . In order to use the computation of the RKHS norm derived in
Appendix D.2, we need to make sure that g has the proper parity. This can de done by
removing all harmonics with k 6 s (note that these harmonics are also functions of w> x,
and thus the function that we obtain is also a function of w> x). That function then has a
squared RKHS norm equal to
X
kgk k2L2 (Sd ) λ−2
k .
k>s,λk 6=0
The summand has an asymptotic equivalent proportional to k −d−2s−1 k d−1 k d+2α+1 which is
equal to k d+2α−2s−1 . Thus if d + 2α − 2s > 0, the series is divergent (the function is not in
the RKHS), i.e., if s 6 α + d2 .
45
Bach
Best approximation by a function in G2 . The squared norm of the k-th harmonic

kgk k2L2 (Sd ) goes down to zero as k −2s−2 and the squared RKHS norm of a h is equivalent
X
to khk k2L2 (Sd ) k d+2α+1 . Given δ, we may then find the function h such that γ2 (h)2 =
P k>0 2 d+2α+1 6 δ 2 with smallest L (Sd ) norm distance to g, that is,
P
k>0 khk kL2 (Sd ) k 2 k>0 kgk −
2
P
hk kL2 (Sd ) . The optimal approximation is hk = αk gk for some αk ∈ R+ , with error k>0 (1 −
αk )2 kgk k2L2 (Sd ) ∼ 2 −2s−2 and squared γ -norm 2 d+2α+1 k −2s−2 =
P P
k>0 (1 − αk ) k 2 k>0 αk k
2 d+2α−2s−1 . The optimal α is obtained by considering a Lagrange multiplier λ such
P
k>0 αk k k
that (αk − 1)k −2s−2 + λαk k d+2α−2s−1 = 0, that is, αk = (k −2s−2 + λk d+2α−2s−1 )−1 k −2s−2 =
(1 + λk d+2α+1 −1
) . We then have
X X
αk2 k d+2α−2s−1 = (1 + λk d+2α+1 )−2 k d+2α−2s−1
k>0 k>0
Z ∞
≈ (1 + λtd+2α+1 )−2 td+2α−2s−1 dt by approximation by an integral,
0
Z ∞
∝ (1 + u)−2 d(td+2α−2s ) with the change of variable u = λtd+2α+1
0
−(d+2α−2s)/(d+2α+1)
∝ λ up to constants,
which should be of order δ 2 (this gives the scaling of λ as a function of δ). Then the squared
error is
X X λ2 t2d+4α+2
(1 − αk )2 k −2s−2 = k −2s−2
(1 + λk d+2α+1 )2
k>0 k>0
∞
λ2 t2d+4α−2s
Z
≈ dt
0 (1 + λtd+2α+1 )2
2 −(2d+4α−2s+1)/(d+2α+1)
≈ λ λ = λ−(2d+4α−2s+1−2d−4α−2)/(d+2α+1)
= λ(2s+1)/(d+2α+1) ≈ δ −2(2s+1)/(d+2α−2s) ,
and thus the (non-squared) approximation error scales as δ −(2s+1)/(d+2α−2s) . For s = 1, this
leads to a scaling as δ −3/(d+2α−2) .
D.6 Proof of Prop. 4

For α = 1, by writing v > x = (v > x)+ − (−v > x)+ we obtain the upperbound γ1 (g) 6 2. For
all other situations, we may compute
X kgk k2L2 (Sd )
γ2 (g)2 = .
λ2k
k>0
For g a linearRfunction gk = 0 exceptRfor k = 1, for which, we have g1 (x) = v > x, and thus
kgk k2L2 (Sd ) = Sd (v > x)2 dτd (x) = v > ( Sd xx> τd (x))v = 1. This implies that γ2 (g) = λ−1
1 .
d−1 α Γ(α/2)Γ(d/2+1)
Given the expression (from Appendix D.2) λ1 = d 4π Γ(α/2+d/2+1) for α > 1 and λ1 =
d−1
2dπ .
46
Appendix E. Computing `2 -Haussdorff Distance between Ellipsoids

We assume that we are given two ellipsoids defined as (x − a)> A−1 (x − a) 6 1 and (x −
b)> B −1 (x − b) 6 1 and we want to compute their Hausdorff distance. This leads to the two
equivalent problems
max w> (a − b) − kB 1/2 wk2 + kA1/2 wk2 ,

kwk2 61
max min ka + A1/2 u − b − B 1/2 vk2 ,

kuk2 61 kvk2 61
which are related by w = a + A1/2 u − b − B 1/2 v. We first review classical methods for
optimization of quadratic functions over the `2 -unit ball.
Minimizing convex quadratic forms over the sphere. We consider the following
convex optimization problem, with Q < 0; we have by Lagrangian duality:
1 >
min x Qx − q > x
kxk2 61 2
1 λ
max min x> Qx − q > x + (kxk22 − 1)
λ>0 x∈Rd 2 2
1 λ
max − q > (Q + λI)−1 q − with x = (Q + λI)−1 q.
λ>0 2 2
If kQ−1 qk2 6 1, then λ = 0 and x = Q−1 q. Otherwise, at the optimum, λ > 0 and
1
kxk22 = q > (Q + λI)−2 q = 1, which implies 1 6 λ+λmin > −1
(Q) q Q q, which leads to λ 6
q > Q−1 q − λmin (Q), which is important to reduce the interval of possible λ. The optimal λ
may then be obtained by binary search (from a single SVD of Q).
Minimizing concave quadratic forms over the sphere. We consider the following
non-convex optimization problem, with Q < 0, for which strong Lagrangian duality is known
to hold (Boyd and Vandenberghe, 2004):
1 1
min − x> Qx + q > x = min − x> Qx + q > x
kxk2 61 2 kxk2 =1 2
1 λ
max min − x> Qx + q > x + (kxk22 − 1)
λ>0 x∈Rd 2 2
1 λ
max − q > (λI − Q)−1 q − with x = (Q − λI)−1 q.
λ>λmax (Q) 2 2
1
At the optimum, we have q > (λI − Q)−2 q = 1, which implies 1 6 [λ−λmax (Q)]2
kqk22 , which
leads to 0 6 λ − λmax (Q) 6 kqk2 . We may perform binary search on λ from a single SVD
of Q.
Computing the Haussdorff distance. We need to compute:
1
max min ka + A1/2 u − b − B 1/2 vk22
kuk2 61 kvk2 61 2
47
Bach
1 λ
= max max min ka + A1/2 u − b − B 1/2 vk22 + (kvk22 − 1)
kuk2 61 λ>0 kvk2 61 2 2
λ 1 1
= max max − + ka − b + A1/2 uk2 − (a − b + A1/2 u)> B(B + λI)−1 (a − b + A1/2 u)
kuk2 61 λ>0 2 2 2
λ λ
= max max − + (a − b + A1/2 u)> (B + λI)−1 (a − b + A1/2 u)
kuk2 61 λ>0 2 2
with v = (B + λI)−1 B 1/2 (a − b + A1/2 u). The interval in λ which is sufficient to explore is
λ ∈ [0, −λmin (B) + (ka − bk22 + λmax (A1/2 ))2 ],
which are bounds that are independent of u.

Given λ > 0, we have the problem of
λ µ λ
min max (a − b + A1/2 u)> (B + λI)−1 (a − b + A1/2 u) − (kuk22 − 1) −
µ>0 u∈Rd 2 2 2
λ µ−λ
= min max (a − b)> (B + λI)−1 (a − b) + + λu> A1/2 (B + λI)−1 (a − b)
µ>0 u∈Rd 2 2
1
− u> (µI − λA1/2 (B + λI)−1 A1/2 )u
2
λ µ−λ
= min (a − b)> (B + λI)−1 (a − b) +
µ>0 2 2
+λ2 (a − b)> (B + λI)−1 A1/2 (µI − λA1/2 (B + λI)−1 A1/2 )−1 A1/2 (B + λI)−1 (a − b)
µ
We have u = ( I − A1/2 (B + λI)−1 A1/2 )−1 A1/2 (B + λI)−1 (a − b), leading to w ∝ (λ−1 B −
λ
µ−1 A + I)(a − b). We need µλ > λmax (A1/2 (B + λI)−1 A1/2 ). Moreover
µ
06 − λmax (A1/2 (B + λI)−1 A1/2 ) 6 kA1/2 (B + λI)−1 (a − b)k.
λ
This means that the `2 -Haussdorff distance may be computed by solving in λ and µ, by
exhaustive search with respect to λ and by binary search (or Newton’s method) for µ. The
complexity of each iteration is that of a singular value decomposition, that is O(d3 ). For
more details on optimization of quadratic functions on the unit-sphere, see Forsythe and
Golub (1965).
References
P.-A. Absil, R. Mahony, and R. Sepulchre. Optimization Algorithms on Matrix Manifolds.
Princeton University Press, 2009.
R. A. Adams and J. F. Fournier. Sobolev Spaces, volume 140. Academic Press, 2003.
K. Atkinson and W. Han. Spherical Harmonics and Approximations on the Unit Sphere:
an Introduction, volume 2044. Springer, 2012.
48
F. Bach. Consistency of the group Lasso and multiple kernel learning. Journal of Machine
Learning Research, 9:1179–1225, 2008a.
F. Bach. Exploring large feature spaces with hierarchical multiple kernel learning. In
Advances in Neural Information Processing Systems (NIPS), 2008b.
F. Bach. Convex relaxations of structured matrix factorizations. Technical Report 00861118,

HAL, 2013.
F. Bach. Duality between subgradient and conditional gradient methods. SIAM Journal
on Optimization, 25(1):115–129, 2015.
F. Bach. On the equivalence between kernel quadrature rules and random feature expan-
sions. Journal of Machine Learning Research, 18:1–38, 2017.
A. R. Barron. Universal approximation bounds for superpositions of a sigmoidal function.

IEEE Transactions on Information Theory, 39(3):930–945, 1993.
P. L. Bartlett and S. Mendelson. Rademacher and Gaussian complexities: Risk bounds and
structural results. Journal of Machine Learning Research, 3:463–482, 2003.
P. L. Bartlett, M. I. Jordan, and J. D. McAuliffe. Convexity, classification, and risk bounds.

Journal of the American Statistical Association, 101(473):138–156, 2006.
A. Barvinok. A Course in Convexity, volume 54. American Mathematical Society, 2002.
Y. Bengio, N. Le Roux, P. Vincent, O. Delalleau, and P. Marcotte. Convex neural networks.

In Advances in Neural Information Processing Systems (NIPS), 2006.
A. Berlinet and C. Thomas-Agnan. Reproducing Kernel Hilbert Spaces in Probability and

Statistics, volume 3. Springer, 2004.
E. D. Bolker. A class of convex bodies. Transactions of the American Mathematical Society,

145:323–345, 1969.
L. Bottou and O. Bousquet. The tradeoffs of large scale learning. In Advances in Neural
Information Processing Systems (NIPS), 2008.
J. Bourgain and J. Lindenstrauss. Projection bodies. In Geometric Aspects of Functional

Analysis, pages 250–270. Springer, 1988.
J. Bourgain, J. Lindenstrauss, and V. Milman. Approximation of zonoids by zonotopes.

Acta Mathematica, 162(1):73–141, 1989.
S. P. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
L. Breiman. Hinging hyperplanes for regression, classification, and function approximation.

IEEE Transactions on Information Theory, 39(3):999–1013, 1993.
P. Bühlmann and S. Van De Geer. Statistics for high-dimensional data: methods, theory
and applications. Springer, 2011.
49
Bach
M. Burger and A. Neubauer. Error bounds for approximation with neural networks. Journal
of Approximation Theory, 112(2):235–250, 2001.
Y. Cho and L. K. Saul. Kernel methods for deep learning. In Advances in Neural Information
Processing Systems (NIPS), 2009.
A. S. Dalalyan, A. Juditsky, and V. Spokoiny. A new algorithm for estimating the effective
dimension-reduction subspace. Journal of Machine Learning Research, 9:1647–1678, 2008.
V. F. Dem’yanov and A. M. Rubinov. The minimization of a smooth convex functional on

a convex set. SIAM Journal on Control, 5(2):280–294, 1967.
R. A. DeVore, R. Howard, and C. Micchelli. Optimal nonlinear approximation. Manuscripta

Mathematica, 63(4):469–478, 1989.
M. Dudik, Z. Harchaoui, and J. Malick. Lifted coordinate descent for learning with trace-
norm regularization. In Proceedings of the International Conference on Artificial Intelli-
gence and Statistics (AISTATS), 2012.
J. C. Dunn and S. Harshbarger. Conditional gradient algorithms with open loop step size
rules. Journal of Mathematical Analysis and Applications, 62(2):432–444, 1978.
H. Edelsbrunner. Algorithms in Combinatorial Geometry, volume 10. Springer, 1987.
L. C. Evans and R. F. Gariepy. Measure Theory and Fine Properties of Functions, volume 5.
CRC Press, 1991.
G. E. Forsythe and G. H. Golub. On the stationary values of a second-degree polynomial

on the unit sphere. Journal of the Society for Industrial & Applied Mathematics, 13(4):
1050–1068, 1965.
M. Frank and P. Wolfe. An algorithm for quadratic programming. Naval Research Logistics
Quarterly, 3(1-2):95–110, 1956.
J. H. Friedman and W. Stuetzle. Projection pursuit regression. Journal of the American

Statistical Association, 76(376):817–823, 1981.
C. Frye and C. J. Efthimiou. Spherical Harmonics in p Dimensions. Technical Report

1205.3548, ArXiv, 2012.
K. Fukumizu, F. Bach, and M. I. Jordan. Dimensionality reduction for supervised learning

with reproducing kernel hilbert spaces. Journal of Machine Learning Research, 5:73–99,
2004.
C. Gu. Smoothing Spline ANOVA Models, volume 297. Springer, 2013.
L. J. Guibas, A. Nguyen, and L. Zhang. Zonotopes as bounding volumes. In Proceedings of

the ACM-SIAM symposium on Discrete Algorithms, 2003.
V. Guruswami and P. Raghavendra. Hardness of learning halfspaces with noise. SIAM

Journal on Computing, 39(2):742–765, 2009.
50
L. Györfi and A. Krzyzak. A Distribution-free Theory of Nonparametric Regression.

Springer, 2002.
Z. Harchaoui, A. Juditsky, and A. Nemirovski. Conditional gradient algorithms for norm-
regularized smooth convex optimization. Mathematical Programming, pages 1–38, 2013.
T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Springer,
2009. 2nd edition.
T. J. Hastie and R. J. Tibshirani. Generalized Additive Models. Chapman & Hall, 1990.
S. Haykin. Neural Networks: A Comprehensive Foundation. Prentice Hall, 1994.
G. E. Hinton and Z. Ghahramani. Generative models for discovering sparse distributed
representations. Philosophical Transactions of the Royal Society of London. Series B:
Biological Sciences, 352(1358):1177–1190, 1997.
M. Jaggi. Revisiting Frank-Wolfe: Projection-free sparse convex optimization. In Proceed-
ings of the International Conference on Machine Learning (ICML), 2013.
S. M. Kakade, K. Sridharan, and A. Tewari. On the complexity of linear prediction: Risk
bounds, margin bounds, and regularization. In Advances in Neural Information Process-
ing Systems (NIPS), 2009.
A. R. Klivans and A. A. Sherstov. Cryptographic hardness for learning intersections of
halfspaces. In Annual Symposium on Foundations of Computer Science (FOCS), 2006.
V. Koltchinskii. Rademacher penalties and structural risk minimization. IEEE Transactions
on Information Theory, 47(5):1902–1914, 2001.
S. König. Computational aspects of the Hausdorff distance in unbounded dimension. Tech-
nical Report 1401.1434, ArXiv, 2014.
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolu-
tional neural networks. In Advances in Neural Information Processing Systems, 2012.
V. Kurkova and M. Sanguineti. Bounds on rates of variable-basis and neural-network
approximation. IEEE Transactions on Information Theory, 47(6):2659–2665, Sep 2001.
G. Lan. The complexity of large-scale convex programming under a linear optimization
oracle. Technical Report 1309.5550, arXiv, 2013.
N. Le Roux and Y. Bengio. Continuous neural networks. In Proceedings of the International
Conference on Artificial Intelligence and Statistics (AISTATS), 2007.
W. S. Lee, P. L. Bartlett, and R. C. Williamson. Efficient agnostic learning of neural
networks with bounded fan-in. IEEE Transactions on Information Theory, 42(6):2118–
2132, 1996.
M. Leshno, V. Y. Lin, A. Pinkus, and S. Schocken. Multilayer feedforward networks with
a nonpolynomial activation function can approximate any function. Neural Networks, 6
(6):861–867, 1993.
51
Bach
K.-C. Li. Sliced inverse regression for dimension reduction. Journal of the American Sta-
tistical Association, 86(414):316–327, 1991.
Y. Lin and H. H. Zhang. Component selection and smoothing in multivariate nonparametric

regression. Annals of Statistics, 34(5):2272–2297, 2006.
Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of
training neural networks. In Advances in Neural Information Processing Systems, 2014.
V. Maiorov. Approximation by neural networks and learning theory. Journal of Complexity,

22(1):102–117, 2006.
V. E. Maiorov and R. Meir. On the near optimality of the stochastic approximation of

smooth functions by neural networks. Advances in Computational Mathematics, 13(1):
79–103, 2000.
Y. Makovoz. Uniform approximation by neural networks. Journal of Approximation Theory,

95(2):215–228, 1998.
J. Matoušek. Improved upper bounds for approximation by zonotopes. Acta Mathematica,

177(1):55–73, 1996.
H. N. Mhaskar. On the tractability of multivariate integration and approximation by neural

networks. Journal of Complexity, 20(4):561–590, 2004.
H. N. Mhaskar. Weighted quadrature formulas and approximation by zonal function net-

works on the sphere. Journal of Complexity, 22(3):348–370, 2006.
G. F. Montufar, R. Pascanu, K. Cho, and Y. Bengio. On the number of linear regions of

deep neural networks. In Advances in neural information processing systems, 2014.
V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines.
In Proceedings of International Conference on Machine Learning (ICML), 2010.
R. M. Neal. Bayesian Learning for Neural Networks. PhD thesis, University of Toronto,
1995.
Y. Nesterov. Semidefinite relaxation and nonconvex quadratic optimization. Optimization

Methods and Software, 9(1-3):141–160, 1998.
Y. Nesterov. Introductory lectures on convex optimization: a basic course. Kluwer Academic

Publishers, 2004.
Y. Nesterov. Smooth minimization of non-smooth functions. Mathematical Programming,

103(1):127–152, 2005.
P. P. Petrushev. Approximation by ridge functions and neural networks. SIAM Journal on

Mathematical Analysis, 30(1):155–189, 1998.
A. Pinkus. Approximation theory of the MLP model in neural networks. Acta Numerica,
8:143–195, 1999.
52
A. Rahimi and B. Recht. Random features for large-scale kernel machines. In Advances in
Neural Information Processing Systems (NIPS), 2007.
P. Ravikumar, H. Liu, J. Lafferty, and L. Wasserman. SpAM: Sparse additive models. In

Advances in Neural Information Processing Systems (NIPS), 2008.
R. T. Rockafellar. Convex Analysis. Princeton University Press, 1997.
F. Rosenblatt. The perceptron: a probabilistic model for information storage and organi-
zation in the brain. Psychological Review, 65(6):386, 1958.
S. Rosset, J. Zhu, and T. Hastie. Boosting as a regularized path to a maximum margin

classifier. Journal of Machine Learning Research, 5:941–973, 2004.
S. Rosset, G. Swirszcz, N. Srebro, and J. Zhu. `1 -regularization in infinite dimensional

feature spaces. In Proceedings of the Conference on Learning Theory (COLT), 2007.
W. Rudin. Real and Complex Analysis. Tata McGraw-Hill Education, 1987.
D. E. Rumelhart, G. E Hinton, and R. J. Williams. Learning representations by back-

propagating errors. Nature, 323:533–536, 1986.
R. Schneider. Zu einem problem von shephard über die projektionen konvexer körper.
Mathematische Zeitschrift, 101(1):71–82, 1967.
S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning: From Theory to

Algorithms. Cambridge University Press, 2014.
J. Shawe-Taylor and N. Cristianini. Kernel Methods for Pattern Analysis. Cambridge

University Press, 2004.
A. J. Smola, Z. L. Ovari, and R. C. Williamson. Regularization with dot-product kernels.

Advances in Neural Information Processing Systems (NIPS), 2001.
K. Sridharan. Learning from an Optimization Viewpoint. PhD thesis, Toyota Technological

Institute at Chicago, 2012.
U. von Luxburg and O. Bousquet. Distance–based classification with Lipschitz functions.

Journal of Machine Learning Research, 5:669–695, 2004.
H. Whitney. Analytic extensions of differentiable functions defined in closed sets. Transac-

tions of the American Mathematical Society, 36(1):63–89, 1934.
M. Yuan and Y. Lin. Model selection and estimation in regression with grouped variables.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1):49–67,
2006.
X. Zhang, D. Schuurmans, and Y. Yu. Accelerated training for matrix-norm regularization:

A boosting approach. In Advances in Neural Information Processing Systems (NIPS),
2012.
53

Breaking The Curse of Dimensionality With Convex Neural Networks

Uploaded by

Copyright:

Available Formats

Breaking The Curse of Dimensionality With Convex Neural Networks

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Breaking The Curse of Dimensionality With Convex Neural Networks

Uploaded by

Copyright:

Available Formats

Journal of Machine Learning Research 18 (2017) 1-53 Submitted 12/14; Revised 10/16; Published 4/17

Breaking the Curse of Dimensionality

Francis Bach [email protected]

Editor: Yoshua Bengio

σ(u) = max{u, 0}α = (u)α+ ,

– In Section 4, we provide a detailed theoretical analysis of the approximation prop-

– In Section 5, we study the generalization properties under a standard supervised learn-

2. Learning from Continuously Infinitely Many Basis Functions

2.1 Variation Norm

where µ is a signed Radon measure on V with finite total variation |µ|(V).

2.2 Representation from Finitely Many Functions

2.3 Corresponding Reproducing Kernel Hilbert Space (RKHS)

2.4 Supervised Machine Learning

f¯t = arg min hJ ′ (ft ), f i

Figure 1: Conditional gradient algorithm for minimizing a smooth functional J on F1δ =

Approximation error vs. estimation error. We consider an ε-approximate minimizer

2.5 Incremental Conditional Gradient Algorithms

Conditional gradient algorithm. We assume the functional J is convex and L-smooth,

ft+1 = (1 − ρt )ft + ρt f¯t .

hfˆ, gi 6 max hf, gi 6 κ hfˆ, gi,

for a fixed κ > 1. In Appendix B, we propose a modification of the conditional gradient

Approximation of any function by a finite number of basis functions. The Frank-

3. Neural Networks with Non-decreasing Positively Homogeneous

for specific activation functions σ. We assume that σ is non-decreasing and positively

ϕv (x) = σ(w> x + b) = σ(v > z) = (v > z)α+ ,

which now becomes a function of z ∈ Rd+1 .

– Conversely, given f ∈ F1 , for z = (t> , a)> ∈ Sd , we define g(z) = g(t, a) = f ( Rt α

Figure 2: Sending a ball to a spherical cap.

3.1 Corresponding Positive-definite Kernels

kα (x, x0 ) = E[(w> x + b)α+ (w> x0 + b)α+ ],

kxk22 kx0 k22

3.2 Incremental Optimization Problem for α = 0

3.3 Incremental Optimization Problem for α = 1

= max max min v [T+> b+ − T−> b− ]

= max min max v > [T+> b+ − T−> b− ] by Fenchel duality,

= max min kT+> b+ − T−> b− kq ,

where T+ ∈ Rn+ ×d has rows ti , i ∈ I+ and T− ∈ Rn− ×d has rows ti , i ∈ I− , with v ∈

max min kT+> b+ − T−> b− kq .

for some vectors ti , i = 1, . . . , r (Bolker, 1969). See an illustration in Figure 3. They

Approximation by ellipsoids. Centrally symmetric convex polytopes (w.l.o.g. centered

3.4 Incremental Optimization Problem for α > 2

where T+ ∈ Rn+ ×d has rows ti , i ∈ I+ and T− ∈ Rn− ×d has rows ti , i ∈ I− , with v ∈

4.1 Approximation by a Finite Number of Basis Functions

which is a decomposition of g as a difference of positively homogenous convex functions.

µ+ (v) = ri=1 ηi δ(v −vi )

Proposition 1 (Number of units - α = 1) Let ε ∈ (0, 1/2). For any function g in G1 ,

Proof Without loss of generality, we assume γ(g) = 1. It is shown by Matoušek (1996)

µ+ (Sd ) dµ+ (v) µ− (Sd ) dµ− (v)

4.2 Sufficient Conditions for Finite Variation

Proposition 2 (Finite variation on the sphere) Assume that g : Sd → R is such that

We can make the following observations:

– Dependence on d: a key feature of the sufficient condition is the dependence on d, that

– Special case d = 1, α = 0: differentiable functions on the sphere in R2 , with bounded

4.3 Approximation of Lipschitz-continuous Functions

Proposition 3 (Approximation of Lipschitz-continuous functions on the sphere)

– Dependence on d and α: the rate of approximation is increasing in d and α. In

– Special case d = 1 and α = 0: up to the logarithmic term we recover the result of

4.4 Linear Functions