Inverse Discrete complex-to-real Fourier Transformation (IRDFT)¶

Versioned name : IRDFT-9

Category : Signal processing

Short description : IRDFT operation performs the inverse complex-to-real discrete Fourier transformation of the input tensor by specified dimensions.

Attributes :

No attributes available.


Inputs

• 1 : data - Input tensor of type T with data for the IRDFT transformation. The last dimension of the input tensor must be equal to 2, that is the input tensor shape must have the form [D_0, D_1, ..., D_{N-1}, 2], representing the real and imaginary components of complex numbers in [:, ..., :, 0] and in [:, ..., :, 1] correspondingly. Required.

• 2 : axes - 1D tensor of type T_IND specifying dimension indices where IRDFT is applied, and axes is any unordered list of indices of different dimensions of the input tensor, for example, [0, 4], [4, 0], [4, 2, 1], [1, 2, 3], [-3, 0, -2]. These indices should be integers from -(r - 1) to (r - 2) inclusively, where r = rank(data). A negative axis a is interpreted as an axis r - 1 + a. Other dimensions do not change. The order of elements in the axes attribute matters, and is mapped directly to elements in the third input signal_size. Required.

• NOTE : The following constraint must be satisfied: rank(data) >= len(axes) + 1 and (rank(data) - 1) not in axes and (-1) not in axes.

• 3 : signal_size - 1D tensor of type T_SIZE describing signal size with respect to axes from the input axes. If signal_size[i] == -1, then IRDFT is calculated for full size of the axis axes[i]. If signal_size[i] > data_shape[: r - 1][axes[i]], then input data is zero-padded with respect to the axis axes[i] at the end. Finally, if signal_size[i] < data_shape[: r - 1][axes[i]], then input data is trimmed with respect to the axis axes[i]. More precisely, if signal_size[i] < data_shape[: r - 1][axes[i]], the slice 0: signal_size[i] of the axis axes[i] is considered. Optionally, with default value [data_shape[: r - 1][a] for a in axes].

• NOTE : If the input signal_size is specified, then the size of signal_size must be the same as the size of axes.

Outputs

• 1 : Resulting tensor with elements of the same type as input data tensor and with rank r - 1, where r = rank(data). The shape of the output has the form [S_0, S_1, ..., S_{r-2}], where all S_a are calculated as follows:

1. Calculate normalized_axes, where each normalized_axes[i] = axes[i], if axes[i] >= 0, and normalized_axes[i] = axes[i] + r - 1 otherwise.

2. If a not in normalized_axes, then S_a = data_shape[a].

3. If a in normalized_axes, then a = normalized_axes[i] for some i. In such case, S_a = 2 \* (data_shape[a] - 1) if the signal_size input is not specified, or, if it is specified, signal_size[i] = -1; and S_a = signal_size[a] otherwise.

• When i != len(normalized_axes) - 1, S_a is calculated as S_a = data_shape[a] if the signal_size input is not specified, or, if it is specified, signal_size[i] = -1; and S_a = signal_size[a] otherwise.

• When i = len(normalized_axes) - 1, S_a is calculated as S_a = 2 \* (data_shape[a] - 1) if the signal_size input is not specified, or, if it is specified, signal_size[i] = -1; and S_a = signal_size[a] otherwise.

Types

• T : any supported floating-point type.

• T_IND : int64 or int32.

• T_SIZE : int64 or int32.

Detailed description : IRDFT performs the discrete Fourier transformation of the input tensor, according to the following rules.

For simplicity, assume that an input tensor A has the shape [B_0, ..., B_{k-1}, M_0, ..., M_{q-1}, 2], axes=[k,...,k + q - 1], and signal_size=[S_0,...,S_{q-1}].

Let D be a value of the input tensor A.

Next, put

$X[j_0,\dots,j_{k-1},j_k,\dots,j_{k+q-1}]=D[j_0,\dots,j_{k-1},j_k,\dots,j_{k+q-1},0]+iD[j_0,\dots,j_{k-1},j_k,\dots,j_{k+q-1},1]$

for all indices j_0,...,j_{k+q-1}, where i is an imaginary unit, that is X is a complex tensor.

Define the complex tensor F with the shape [B_0, ..., B_{k-1}, 2 \* (M_0 - 1), ..., 2 \* (M_{q-1} - 1)] using the formula

$\begin{split}F[j_0,\dots,j_{k-1},j_k,\dots,j_p,\dots,j_{k+q-1}] = \begin{cases}X[j_0,\dots,j_{k-1},j_k,\dots,j_p,\dots,j_{k+q-1}],\text{ when }j_p=0,\dots,M_p-1;\\ \overline{X[j_0,\dots,j_{k-1},j_k,\dots,2(M_{p-1} - 1) - j_p,\dots,j_{k+q-1}]},\text{ otherwise.}\end{cases}\end{split}$

Construct the complex tensor G with the shape [B_0, ..., B_{k-1}, S_0, ..., S_{q-1}] by the following way. If S_a > 2 \* (M_a - 1), then the axis k + a of F will be padded by zeros; if S_a < 2 \* (M_a - 1), then the axis k + a of F will be trimmed, that is, we will consider only the slice 0: S_a of this axis; finally, if S_a = 2 \* (M_a - 1), then we consider the full axis k + a of F.

Let Y be a complex tensor with the shape [B_0, ..., B_{k-1}, S_0, ..., S_{q-1}] such that

$Y[n_0,\dots,n_{k-1},m_0,\dots,m_{q-1}]=\frac{1}{\prod\limits_{j=0}^{q-1}S_j}\sum\limits_{p_0=0}^{S_0}\cdots\sum\limits_{p_{q-1}=0}^{S_{q-1}}X[n_0,\dots,n_{k-1},j_0,\dots,j_{q-1}]\exp\left(2\pi i\sum\limits_{b=0}^{q-1}\frac{m_bj_b}{S_b}\right)$

for all indices n_0,...,n_{k-1}, m_0,...,m_{q-1}.

Finally, the result of the inverse discrete complex-to-real Fourier transform is a real part of the tensor Y.

Calculations for the generic case of axes and signal sizes are similar.

Example :

There is no signal_size input (4D input tensor):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>1</dim>
<dim>161</dim>
<dim>161</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>2</dim> <!-- [1, 2] -->
</port>
<output>
<port id="2">
<dim>1</dim>
<dim>161</dim>
<dim>320</dim>
</port>
</output>
</layer>

There is no signal_size input (3D input tensor):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>161</dim>
<dim>161</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>2</dim> <!-- [0, 1] -->
</port>
<output>
<port id="2">
<dim>161</dim>
<dim>320</dim>
</port>
</output>
</layer>

There is signal_size input (4D input tensor):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>1</dim>
<dim>161</dim>
<dim>161</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>2</dim> <!-- [1, 2] -->
</port>
<port id="2">
<dim>2</dim> <!-- [512, 100] -->
</port>
<output>
<port id="3">
<dim>1</dim>
<dim>512</dim>
<dim>100</dim>
</port>
</output>
</layer>

There is signal_size input (3D input tensor):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>161</dim>
<dim>161</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>2</dim> <!-- [0, 1] -->
</port>
<port id="2">
<dim>2</dim> <!-- [512, 100] -->
</port>
<output>
<port id="3">
<dim>512</dim>
<dim>100</dim>
</port>
</output>
</layer>

There is signal_size input (5D input tensor, -1 in signal_size, unsorted axes):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>16</dim>
<dim>768</dim>
<dim>580</dim>
<dim>320</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>3</dim> <!-- axes input contains  [3, 1, 2] -->
</port>
<port id="2">
<dim>3</dim> <!-- signal_size input contains [170, -1, 1024] -->
</port>
<output>
<port id="3">
<dim>16</dim>
<dim>768</dim>
<dim>1024</dim>
<dim>170</dim>
</port>
</output>
</layer>

There is signal_size input (5D input tensor, -1 in signal_size, unsorted axes, the second example):

<layer ... type="IRDFT" ... >
<input>
<port id="0">
<dim>16</dim>
<dim>768</dim>
<dim>580</dim>
<dim>320</dim>
<dim>2</dim>
</port>
<port id="1">
<dim>3</dim> <!-- axes input contains  [3, 0, 2] -->
</port>
<port id="2">
<dim>3</dim> <!-- signal_size input contains [258, -1, 2056] -->
</port>
<output>
<port id="3">
<dim>16</dim>
<dim>768</dim>
<dim>2056</dim>
<dim>258</dim>
</port>
</output>
</layer>